On October 21, 2026, a compromised LiteLLM 1.107.0 release was published to the Python Package Index (PyPI) and distributed through official container images. The malicious package added files that Python could execute during normal interpreter initialization. An application did not need to import LiteLLM for the payload to run: starting Python in an affected environment could be enough.

Microsoft analyzed two execution paths in the affected wheels: a credential collector launched through a path extension, and a Unix-oriented loader persisted through site-specific configuration. The collector targeted authentication material across developer tools, AI applications, cloud environments, browsers, and infrastructure. A separate database module attempted to read selected credential and model-configuration records from the LiteLLM proxy.

Endpoint observations on Windows and macOS matched the recovered collector’s discovery, encryption, and outbound-connection behavior. In one Windows example, a routine Python invocation to install an unrelated dependency activated the collector. This distinction matters for response: investigators should not start Python or pip inside a suspect environment to determine whether the malicious package is present.

This blog explains the startup mechanisms, the credentials the payload could reach, and the evidence defenders can use to investigate exposure. It also provides Microsoft Defender detection information, advanced hunting queries, and recovery guidance. Removing the package or a release tag is not sufficient recovery. Organizations must address the affected Python environment, any surviving processes, and credentials accessible to the compromised identity.

Attack overview

LiteLLM is commonly deployed as a gateway between applications and model providers. In that position it may have access to model-provider API keys, database connection strings, workload identities, routing configuration, and tenant policy data. The same package is also used on developer workstations, build agents, test systems, and containers, which widened the set of environments that could be exposed by the compromised release.

According to LiteLLM maintainers, stolen developer GitHub credentials were used to alter the release workflow and redirect it to the lookalike repository Berri-AI/litellm rather than the legitimate BerriAI/litellm repository. A credential associated with another developer account was then used to satisfy a production approval step. Microsoft has not independently determined how the attacker first accessed the upstream environment or obtained those credentials.

The compromised package used Python startup mechanisms to execute malicious code automatically, meaning exposure was not limited to systems actively using a LiteLLM gateway. Developer workstations, continuous integration and continuous delivery/deployment (CI/CD) agents, test environments, containers, and other systems that started Python with the affected package installed could also execute the payload. The malicious code supported credential collection and, on Unix-based systems, retrieval and execution of additional Python code.

1. Compromised package distribution

The compromised LiteLLM 1.107.0 release delivered malicious Python startup files to environments where the package was installed.

2. Automatic payload execution

Python startup processing launched the credential collector without requiring the application to actively use LiteLLM gateway functionality.

3. Credential and sensitive-data collection

The collector targeted developer credentials, cloud credentials, browser data, SSH material, environment variables, and LiteLLM database data accessible to the affected process.

4. Encryption and exfiltration

Collected data was encrypted and transmitted to an attacker-controlled endpoint.

5. Conditional retrieval and execution of additional code

On Unix-based systems, the payload supported retrieving and executing additional Python code, although this activity was not confirmed.

Technical details

Compromised package distribution and automatic payload execution

The compromised wheel packaged cook.pth, cook.py, and sitecustomize.py under litellm-1.107.0.data/purelib/. During installation, these files were relocated into the Python environment’s library directory, outside the normal litellm application directory. This placement enabled execution through Python’s normal site initialization, which processes executable import statements in .pth files and later attempts to import the sitecustomize module.

The first execution path began with the single executable line in cook.pth: “import cook”. This loaded cook.py, whose wrapper Base64-decoded and zlib-decompressed an embedded Python program before executing the recovered source. The program identified itself as PRIVCOLLECT and performed credential collection in a daemon thread.

Before starting the thread, the collector checked the inherited PRIVCOLLECT_ACTIVE environment variable. A nonempty value caused it to return; otherwise, it set the flag, limiting duplicate collection in descendant processes. After an upload returned HTTP 200, it attempted to write /tmp/.pc_once_72129118c59f5636c0bb. These duplicate-suppression gates were not containment controls: startup alone did not prove collection or upload completed, because a short-lived process could exit before the thread finished, while write failures or temporary-storage changes could leave the marker absent.

A second startup mechanism used sitecustomize.py, providing a separate Unix-oriented execution path with an embedded collector and functionality to retrieve and execute additional Python code from a remote host. Remote retrieval and execution through this path were not confirmed.

Together, these mechanisms extended potential exposure beyond active LiteLLM deployments: simply starting Python in an affected environment could activate the malicious code, even when LiteLLM itself was never imported. Because the startup files resided outside the litellm application directory, removing only the application files could leave the malicious startup components behind.

Credential and sensitive-data collection

The recovered collector was designed to gather sensitive authentication material across developer tools, AI applications, cloud environments, browsers, and infrastructure. It runs multiple collection modules sequentially with independent error handling, allowing collection to continue even when individual targets are unavailable or inaccessible.

Encryption and exfiltration

After collection, the payload serializes module results and host metadata as JSON. Each run generates a 32-byte AES key and a 16-byte initialization vector. AES-256-CBC encrypts the collection document; an embedded 4,096-bit RSA public key wraps the AES key using OAEP. The OpenSSL path tries SHA-256 and then SHA-1 for OAEP, and the resulting metadata records the selected digest.

The collector places three members in a gzip-compressed tar archive: data.bin, containing the initialization vector (IV) and encrypted collection; key.bin, containing the wrapped AES key; and hdr.json, containing plaintext envelope metadata. It then attempts an HTTP POST to hxxp://www[.]litellm[.]dev/healthcheck, using application/octet-stream, a fixed browser-like User-Agent, and a body checksum in an X-SHA256 header. The upload routine retries and treats only HTTP 200 as success.

The OpenSSL backend creates temporary directories beginning with .pc_e_, stages input and output files and removes them after use.

Conditional retrieval and execution of additional code

In addition to the credential collector, the compromised package contained a separate Unix-oriented loader in sitecustomize.py, designed to retrieve and execute additional Python code. When Python starts, the loader takes an exclusive lock on /tmp/remote_code_loader.lock so that only one instance runs per host and detaches into the background. Each cycle reads a GitHub-hosted domain.txt file, which acts as a mutable pointer to an attacker-controlled server.

If a valid server address is returned and successfully resolved, the loader is designed to request /codepy from that server, compile the returned content, and execute it as Python code within the affected environment. This gives the attacker a mechanism to change the second-stage payload without publishing another compromised LiteLLM package. This loader is separate from the credential collector delivered directly in the compromised package.

At the time the cited researchers examined domain.txt, it was empty and no second-stage retrieval or execution was confirmed. Because the file could be changed, the loader remained a re-armable remote-code execution mechanism.

Mitigation and protection guidance

Microsoft recommends the following actions for organizations that installed, mirrored, or built images with LiteLLM 1.107.0:

• Find and contain every affected copy. Identify LiteLLM 1.107.0 across installed environments, uv and pip caches, internal package mirrors, lock files, build outputs, and container images. Scope mutable container tags and the relevant exposure windows once validated and compare artifacts with the customer-ready hashes in the indicators of compromise section. For background, see secure software supply chain practices and Docker image digests. Digests are preferable to mutable tags because they identify immutable image content.

• Prevent Python execution and investigate from outside the environment. Isolate affected developer workstations, build agents, test systems, containers, and AI gateway hosts. Where operationally safe, stop running Python processes and prevent automation from restarting them. Do not run python, pip, or other commands that invoke the affected interpreter during evidence collection; inspect the environment with trusted tools from a known-clean system. Python normally processes site configuration, including executable import lines in .pth files, during initialization.

• Rebuild from verified artifacts. Confirm that cook.pth, cook.py, and sitecustomize.py were removed from the environment library path. Rebuilding is more reliable than assuming that deleting the litellm directory cleaned the top-level startup files.

• Rotate reachable credentials from a known-clean machine. Include model-provider keys, cloud and workload credentials, CI/CD secrets, source-control and registry tokens, SSH and GPG keys, AI coding-tool tokens, and browser sessions available to the affected identity. On proxy hosts, include stored provider credentials and DATABASE_URL. Revoke affected identity sessions rather than assuming password or key rotation immediately invalidates every active token or application session. See revoke user access in Microsoft Entra ID.

• Correlate before declaring compromise. Combine file hashes, Python process ancestry, .pc_e_ cryptographic staging, network destination, workload identity, and subsequent cloud or identity use. Do not use package presence or a generic filename as the sole verdict.

• Strengthen dependency controls. Use hash-pinned lock files, verify package and image provenance, use container digests instead of floating tags, and admit packages to internal mirrors only after review. See best practices for a secure software supply chain and dependency scanning in Azure DevOps.

• Restrict outbound access. Developer workstations, build agents, test systems, containers, and AI gateway hosts should not have unrestricted outbound HTTP access unless required. Route outbound traffic through managed egress controls, deny unnecessary destinations, and allowlist required services. See control outbound internet access from Azure.

• Enable Microsoft Defender for Endpoint protections. Onboard supported developer workstations, build agents, and AI gateway servers to Microsoft Defender for Endpoint, and enable Microsoft Defender Antivirus real-time and cloud-delivered protection where supported. See configure always-on protection and configure cloud protection.

Microsoft Defender XDR detections

As of October 22, 2026, the following Microsoft Defender Antivirus signatures and Microsoft Defender for Endpoint alerts were available for activity associated with the compromised LiteLLM 1.107.0 release. Detection names and availability may change as protection is updated.

Threat intelligence reports

Microsoft customers can use the following reports in Microsoft products to get the most up-to-date information about the threat actor, malicious activity, and techniques discussed in this blog. This report provides intelligence, protection information, and recommended actions to prevent, mitigate, or respond to associated threats found in customer environments.

Microsoft Defender XDR Threat analytics

Inside the LiteLLM 1.107.0 supply chain compromise: Credential theft through Python startup hooks

Advanced hunting

These queries support investigation of the stages discussed above. Select the appropriate lookback in Defender XDR and correlate results with environment, process and identity evidence. They are investigation leads, not standalone verdicts.

Connections to the collector domain

Normalize URL/hostname observations before matching the campaign domain. The query rejects suffix lookalikes such as www[.]litellm[.]dev[.]example; a match still needs process and workload context.

DeviceNetworkEvents
| extend UrlForParsing = iff(RemoteUrl contains "://",
    RemoteUrl, strcat("http://", RemoteUrl))
| extend RemoteHost = trim_end(@"\.",
    tolower(tostring(parse_url(UrlForParsing).Host)))
| where RemoteHost in ("www.litellm.dev", "litellm.dev")
| project Timestamp, DeviceId, DeviceName, ActionType,
    RemoteUrl, RemoteIP, RemotePort, InitiatingProcessFileName,
    InitiatingProcessId, InitiatingProcessCreationTime, ReportId

Known startup-file hashes

The four hashes cover LF and CRLF variants. A missing SHA-256 field limits this hunt; a generic filename is not an equivalent match.

DeviceFileEvents
| where SHA256 in~ (
    "9dd0a3926f96a37eb86cf1ee2310d9be93139cc4ebef3ef9e2efe9531390be25",
    "03ac2a8b752605d1d7710979c17ff18b03dd978203ba23ba97fd8ce3d6f897ea",
    "34bef04fe88ac9d0aae0475630b1738681d15b94f233ac8aa46eacf5968305fa",
    "b6122440d0d9cb1eb1ade9933f49214da8c31288de7710c076c2680dfa528465",
    "7f191632e7320705d154b85204f1ae546b543acbd7e8861229b76f5042b92bdb",
    "84dc8c82c0a4a4a62552ef826b49fcc86d1a5c4a5d0f2c06665723f93d129d7d")
| project Timestamp, DeviceId, DeviceName, ActionType,
    FileName, FolderPath, SHA256, InitiatingProcessFileName,
    InitiatingProcessId, InitiatingProcessCreationTime, ReportId

Inline database-collection program

Look for the sensitive table/field pair together with the query API, serialization and complete output-suppression pattern. Legitimate exports may overlap. The recovered parent’s /proc prerequisite limits its applicability; raw arguments are omitted because they can contain a database secret.

DeviceProcessEvents
| where FileName startswith "python" or FileName startswith "platform-python"
| where strlen(ProcessCommandLine) 

Collector-like encryption staging

The staging prefix plus encryption operation identifies a behavioral lead, not an upload or a uniquely malicious OpenSSL invocation. Alternate backends are outside this query, and key-bearing arguments are not projected.

DeviceProcessEvents
| where FileName in~ ("openssl", "openssl.exe")
| where ProcessCommandLine contains_cs ".pc_e_"
| where (ProcessCommandLine contains_cs " enc "
        and ProcessCommandLine contains_cs "-aes-256-cbc")
    or ProcessCommandLine contains_cs " pkeyutl "
    or ProcessCommandLine contains_cs " rsautl "
| extend CryptoOperation = iff(
    ProcessCommandLine contains_cs " enc ", "AES packaging", "RSA wrapping")
| project Timestamp, DeviceId, DeviceName, FileName, CryptoOperation,
    ProcessId, ProcessCreationTime, InitiatingProcessFileName,
    InitiatingProcessId, InitiatingProcessCreationTime, ReportId

Vulnerability/Posture Check | Is the device potentially exposed to the affected Litellm?

Use this query to identify devices potentially exposed to TVM-2026-0003. A match indicates potential exposure, not confirmed exploitation or compromise.

DeviceTvmSoftwareVulnerabilities
| where CveId == "TVM-2026-0003"

CloudProcessEvents query to detect encryption before exfiltration

Identifies OpenSSL public-key encryption using /tmp/.pc_ staging paths, consistent with key wrapping before potential exfiltration.

CloudProcessEvents
| where ProcessName == 'openssl' and ProcessCurrentWorkingDirectory has_any ('/runner/','/app','/tmp/')
    and ProcessCommandLine has_all ('pkeyutl','-encrypt','-pubin') and ProcessCommandLine contains '/tmp/.pc_'

MITRE ATT&CK techniques observed

Microsoft mapped the following techniques based on observed activity and analysis of the recovered malicious components; individual entries note where behavior was not confirmed. For standard industry documentation, refer to the MITRE ATT&CK framework.

Initial Access

• T1195.001 Compromise Software Dependencies and Development Tools | Malicious files shipped in LiteLLM 1.107.0 wheels and images; upstream release abuse reported by the maintainer.

Persistence

• T1546.018 Event Triggered Execution: Python Startup Hooks | cook.pth and sitecustomize.py run during Python site initialization.

Execution

• T1059.006 Command and Scripting Interpreter: Python | The collector runs in the interpreter; an inline Python child performs database queries.

Defense Evasion

• T1027.013 Obfuscated Files or Information: Encrypted/Encoded File | Collector source is embedded as Base64-encoded, zlib-compressed content.
• T1140 Deobfuscate/Decode Files or Information | cook.py decodes and decompresses the collector at runtime.

Credential Access

• T1552.001 Unsecured Credentials: Credentials In Files | Reads authentication and configuration files for developer, AI, and cloud tools.
• T1552.004 Unsecured Credentials: Private Keys | Reads selected SSH, GPG, and TLS key material.
• T1552.005 Unsecured Credentials: Cloud Instance Metadata API | Requests AWS and GCP credentials from metadata services; returned credentials were not confirmed in Microsoft telemetry.
• T1555.003 Credentials from Password Stores: Credentials from Web Browsers | Copies browser databases and supports selected Chromium password recovery on Windows.

Discovery

• T1082 System Information Discovery | systeminfo was launched by Python in an observed Windows sequence.
• T1016 System Network Configuration Discovery | ipconfig, route, arp, and netstat were launched by the same Python process.

Collection

• T1005 Data from Local System | Collects selected local histories, environment variables, and configuration.

Command and Control

• T1102.001 Web Service: Dead Drop Resolver | The loader reads its server address from a specific GitHub-hosted file.
• T1105 Ingress Tool Transfer | The loader can retrieve Python code from /codepy; retrieval and execution were not confirmed.

Indicators of compromise

URLs are defanged. Use precise roles and context rather than blocking shared infrastructure indiscriminately.

References

• LiteLLM, Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised
• Python Software Foundation, Site-specific configuration hook.

Learn More

For the latest security research from the Microsoft Threat Intelligence community, check out the Microsoft Threat Intelligence Blog.

To get notified about new publications and to join discussions on social media, follow us on LinkedIn, X (formerly Twitter), and Bluesky.

To hear stories and insights from the Microsoft Threat Intelligence community about the ever-evolving threat landscape, listen to the Microsoft Threat Intelligence podcast.

Review our documentation to learn more about our real-time protection capabilities and see how to enable them within your organization.  

• Learn more about securing Copilot Studio agents with Microsoft Defender 
• Evaluate your AI readiness with our latest Zero Trust for AI workshop.
• Microsoft 365 Copilot AI security documentation
• How Microsoft discovers and mitigates evolving attacks against AI guardrails
• Prompt injection protection in Microsoft Defender for Office 365 – official documentation of the prompt injection protection in Microsoft Defender for Office 365.
• How Microsoft discovers and mitigates evolving attacks against AI guardrails | Microsoft Security Blog
• Manipulating AI memory for profit: The rise of AI Recommendation Poisoning | Microsoft Security Blog – a related example of an AI-era technique observed in email traffic
• Defending the inbox against prompt injection attacks | Microsoft Defender for Office 365 Blog – feature announcement introducing prompt injection protection in Microsoft Defender for Office 365.
• Learn how Microsoft is reimagining the SOC for the agentic era with ISOC in Microsoft Defender

The post Inside the LiteLLM 1.107.0 supply chain compromise: Credential theft through Python startup hooks appeared first on Microsoft Security Blog.

As early as September 2026, Storm-0940, a China-state aligned actor, used the agentic AI penetration testing framework ARTEX to automate attacks targeting hundreds of organizations globally across a broad range of industries. Several of these Storm-0940 ARTEX-driven attacks have resulted in the actor obtaining privileged access to an organization’s cloud infrastructure – including multiple instances where the time from initial targeting to privileged access was less than 3 hours. Microsoft assesses that Storm-0940 was responsible for the recently disclosed ConnectWise supply chain compromise of a multi-tenant application.

This activity demonstrates how threat actors are integrating agentic AI into existing offensive security workflows and allows them to scale their operations exponentially. Agents are not constrained by factors affecting humans – they can work around the clock, on weekends, and on holidays and attempt multiple breach paths simultaneously enabling significantly reduced time from selecting a target to achieving objectives, increasing the speed and impact of complex attacks like supply chain compromises.

Storm-0940’s use of ARTEX accelerated the speed of compromised application abuse to authenticate to Microsoft 365 customer environments through applications that had already been granted access. This allowed the threat actor to operate without requiring an interactive user sign-in. In multiple affected environments, Microsoft observed AI-accelerated follow-on activity through Microsoft Graph, including reconnaissance, application credential changes, and attempts to modify privileges that could enable expanded or persistent access. The activity conducted by the threat actor through the usage of ARTEX framework did not result from a compromise of or a vulnerability in Microsoft corporate systems, products, or services.

Although Microsoft identified Storm-0940-initiated authentication and follow-on activity in multiple customer environments, the presence of an affected application or successful threat actor authentication does not necessarily indicate that customer data was accessed or exfiltrated. Microsoft Defender XDR provides detections and hunting queries for suspicious Microsoft Graph activity and activity associated with known threat infrastructure. Customers should review service principal sign-ins, investigate unexpected application credential or privilege changes, and follow ConnectWise’s published containment guidance while preserving relevant evidence for investigation. For a full list of detections please see the detections section.

What’s happening

Storm-0940 is a China-state aligned actor (with no known industry aliases) that has been active since 2021 and targets organizations across multiple industries for data theft. The actor has historically gained initial access using password spraying and exploitation of internet-facing applications, before using compromised accounts to move through targeted environments. In cloud environments, the actor has also established persistence by adding credentials to existing OAuth applications, allowing continued access to Microsoft 365 resources. The actor has used these applications to access and exfiltrate email and other data through Microsoft Graph.

As early as September 2026, Storm-0940 used the agentic AI penetration testing framework ARTEX to automate attacks targeting hundreds of organizations globally across a broad range of industries. These include aerospace and defense, automotive, chemicals, energy, financial services, government, legal, manufacturing, media, military, semiconductor, software, technology, and think tanks and non-profits. During the ConnectWise supply chain compromise, Storm-0940 used ARTEX to accelerate scanning and reconnaissance, password spraying, and OAuth application abuse at downstream victims.

Several of these Storm-0940 ARTEX-driven attacks have resulted in the actor obtaining privileged access to an organization’s cloud infrastructure – including multiple instances where the time from initial targeting to privileged access was less than 3 hours. When automated with AI, these activities generate a large volume of traffic and log data that may generate alerts. However, the speed of the activity allows very little time for Defenders to respond to those alerts and remediate the actor’s access in a coordinated fashion. Storm-0940’s activity illustrates how a threat actor with a broad target set can use an agentic AI tool and privileged authentication material to quickly attempt to compromise many organizations. While the majority of the targeting wasn’t successful, Storm-0940 achieved privileged cloud persistence at many organizations in a brief time.

What is ARTEX?

ARTEX is an open-source penetration testing framework that can use any AI model to automate a variety of offensive security workflows. The ARTEX framework is a multi-agent LLM system written in GO with a Next.js frontend. While the UI and generated output default to English, the framework can be switched to Korean, Mandarin, Spanish, or Japanese. ARTEX is a system of LLM-driven agents that run an autonomous penetration test from reconnaissance through data exfiltration with little needed human interaction. The agents then accumulate the discovered assets and vulnerabilities into a graph during the penetration test. In October 2026, public reporting noted that the ARTEX framework was being used by threat actors, such as the personal data breaches targeting Korean financial institutions. Microsoft Threat Intelligence has identified threat actors across multiple geographic regions leveraging the ARTEX framework.

ConnectWise published an initial advisory on October 16, 2026, and subsequently expanded its disclosures and containment guidance. The advisory details ConnectWise’s investigation of attempted unauthorized access associated with Cloud Backup. Following the unauthorized access, Storm-0940 used ARTEX to attempt to sign-in to more than 2,250 tenants. While OAuth application abuse is a typical part of Storm-0940’s toolset, in this case the actor used ARTEX to accelerate the activity to machine speed.

Password Spraying

Between October 10 and 14, 2026, Storm-0940 infrastructure at 144.172.64[.]99 conducted an automated cross-tenant password-spray campaign targeting 327 accounts across eight organizations to gain access to Microsoft 365 and Exchange Online. During this large campaign, Microsoft telemetry captured 151,000+ failed logon events, with activity peaking on October 11 at 2,900 failures per minute and concentrating more than 111,000 failures into a single day. The campaign blended broad, low-rate account spraying with sustained, high-volume authentication attempts against selected identities; 48 accounts received more than 1,000 failures, and lockout responses represented nearly 98% of all activity. Despite the scale, no successful authentication, token issuance, Graph call, or directory change was observed.

Downstream customer access

On October 15, immediately after this notable activity, public reports of the compromise of the popular cloud backup application ConnectWise started to surface with direct references to indicators of compromise matching the same network range observed for the earlier password-spray campaign (144.172.64.x). An advisory from the affected cloud backup vendor posted on October 16 shared later additional indicators with one IP address matching exactly the 144.172.64[.]99 infrastructure responsible for the earlier password-spray campaign.

Additionally, after acquiring control of the identities of multiple ConnectWise cloud applications, the AI agents operating on behalf of the threat actor continued to use the same IP range to authenticate to customer environments using the illicit access acquired through these stolen app credentials, effectively taking advantage of the already established trust by these apps within a tenant. In some of these environments, the AI agents (on behalf of the threat actor) progressed beyond authentication to Microsoft Graph and directory operations, including reconnaissance, application credential modifications to establish persistence, and changes involving identities, groups, and privileged roles. Microsoft identified multiple customer environments with evidence of follow-on activity, highlighting how a compromised third-party application with high privileges could provide access beyond its originally intended purpose.

The application credential modification made by Storm-0940’s associated agents in the targeted victims’ environments can lead to persistence through other identities, credentials, and cloud services. These changes can remain active as a backdoor even after the original compromised application is mediated. Organizations should examine service principal sign-ins alongside subsequent Microsoft Graph activity, directory changes, and resource access, correlating suspicious events with authorized administrative activity. The presence of an affected application, or even successful authentication, does not necessarily indicate unauthorized data access or data exfiltration.

Mitigation and protection guidance

Microsoft recommends the following mitigations to reduce the impact of this threat. Check the recommendations card for the deployment status of monitored mitigations.

Immediate actions

• Keep affected ConnectWise application connections disabled until ConnectWise provides formal restoration instructions.
• Do not delete enterprise applications as the first response. Preserve configuration and evidence before making irreversible changes.
• Inventory all four application IDs listed in this profile and record the local service-principal Object ID in each tenant.
• Review every Microsoft 365 tenant where the affected applications are currently or were previously installed or authorized, even if the integration is no longer in use. Search by application ID rather than display name.
• Preserve available Microsoft Entra sign-in and audit records, Microsoft Purview Audit records, SIEM data, and previously configured Microsoft Graph telemetry. Do not delay urgent containment while waiting for exports.
• Review activity starting from at least October 15, 2026. Where records are available, extend service-principal sign-in review to October 1 and investigate earlier related activity, including failed authentications and volume anomalies.
• Review successful and failed service-principal sign-ins separately. Investigate unfamiliar sources and activity outside expected vendor infrastructure; do not limit review to the published IPs.
• Review Microsoft Entra audit records for application or service-principal credential additions; new applications, service principals, or owners; app-role assignments and consent grants; privileged role assignments; and changes to users, groups, authentication methods, or Conditional Access. On confirmed suspicious activity, revoke unnecessary permissions and rotate credentials.
• Review Microsoft 365 audit records to look for reconnaissance activities like listing directories, permissions, and approvals coming from unusual IPs.
• Review Microsoft 365 audit records for unusual mail or file access, downloads, forwarding, inbox or transport rules, delegated mailbox permissions, and related resource activity.
• Preserve evidence, then have an appropriately authorized responder remove confirmed unauthorized credentials, grants, owners, assignments, and identities.

Additional recommendations

• Threat actor misuse of AI accelerates risk across enterprise environments by amplifying existing threats such as exploitation of vulnerabilities, phishing, malware threats, and insider activity. To help organizations stay ahead of AI-enabled threat activity, Microsoft has introduced the Security Dashboard for AI. The dashboard provides users with a unified view of AI security posture by aggregating security, identity, and data risk across Microsoft Defender, Microsoft Entra, and Microsoft Purview. This allows organizations to understand what AI assets exist in their environment, recognize emerging risk patterns, and prioritize governance and security across AI agents, applications, and platforms. To learn more about the Microsoft Security Dashboard for AI see: Assess your organization’s AI risk with Microsoft Security Dashboard for AI (Preview).
• Enable threat protection for AI services in Microsoft Defender for Cloud to identify threats to generative AI applications in real time and for assistance in responding to security issues.
• Apply least privilege to multi-tenant application permissions and consent grants.
• Monitor credential additions, privilege assignments, and anomalous app-only activity, especially to applications with high privileged permissions and scopes.
• Verify that vendor credential rotation and any tenant-level remediation eliminated remaining unauthorized access.

Assess the operational effect on backup and restore workflows and follow current ConnectWise restoration guidance.

Microsoft Defender XDR detections

Microsoft Defender XDR customers can refer to the list of applicable detections below. Microsoft Defender XDR coordinates detection, prevention, investigation, and response across endpoints, identities, email, and apps to provide integrated protection against attacks like the threat discussed in this blog.

Customers with provisioned access can also use Microsoft Security Copilot in Microsoft Defender to investigate and respond to incidents, hunt for threats, and protect their organization with relevant threat intelligence.

Microsoft Defender for Cloud Apps detections

Microsoft Defender for Cloud (Microsoft Defender for Key Vault)

Microsoft Security Copilot

Security Copilot customers can use the standalone experience to create their own prompts or run the following prebuilt promptbooks to automate incident response or investigation tasks related to this threat:

• Incident investigation
• Threat actor profile
• Threat Intelligence 360 report based on MDTI article

Note that some promptbooks require access to plugins for Microsoft products such as Microsoft Defender XDR or Microsoft Sentinel.

Threat intelligence reports

Microsoft customers can use the following reports in Microsoft products to get the most up-to-date information about the threat actor, malicious activity, and techniques discussed in this blog. These reports provide intelligence, protection information, and recommended actions to prevent, mitigate, or respond to associated threats found in customer environments.

Microsoft Defender XDR Threat analytics

Activity Profile: AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise

Indicators of compromise (IOC)

Advanced hunting

NOTE: The following sample queries let you search for a week’s worth of events. To explore up to 30 days’ worth of raw data to inspect events in your network and locate potential indicators related to this campaign for more than a week, go to the Advanced Hunting page > Query tab, select the calendar dropdown menu to update your query to hunt for the Last 30 days.

Microsoft Defender XDR advanced hunting

1- OAuthAppInfo → Is the affected app present, enabled, and what permissions does it hold?

Start with OAuthAppInfo to determine whether any of the affected ConnectWise applications are present in the tenant and understand the access they have been granted. The query identifies the application and its tenant-specific service principal, whether it is enabled, its privilege level, consent information, and the individual permissions assigned to it.

This establishes the customer’s potential exposure and gives investigators the ServicePrincipalId needed for further hunting. A match does not mean the application was abused, only that it is present and potentially had access, so the next step is to determine whether the service principal was actually used. Run query

let ConnectWiseApps = dynamic([
    "506e08f9-6e9a-4da4-942b-a099a03644ae", // ConnectWise Cloud Backup
    "24205314-2b89-4e07-a2a4-579e898a5640", // ConnectWise - MS Teams
    "0de14c77-b9b0-4137-9360-34878a929827" // ConnectWise - MS Graph 
]);
OAuthAppInfo
| where OAuthAppId in (ConnectWiseApps)
| mv-expand Permission = Permissions
| extend
    PermissionName = tostring(Permission.PermissionName),
    TargetApp = tostring(Permission.TargetAppDisplayName),
    TargetAppId = tostring(Permission.TargetAppId),
    PermissionType = tostring(Permission.PermissionType),
    PermissionPrivilegeLevel = tostring(Permission.PrivilegeLevel),
    UsageStatus = tostring(Permission.UsageStatus)
| project
    AppName,
    OAuthAppId,
    ServicePrincipalId,
    AppStatus,
    PrivilegeLevel,
    IsAdminConsented,
    PermissionName,
    PermissionType,
    PermissionPrivilegeLevel,
    TargetApp,
    TargetAppId,
    UsageStatus
| order by AppName asc, PermissionPrivilegeLevel desc

2 – EntraIdSpnSignInEvents + GraphAPIAuditEvents → Did the service principal authenticate, and what Microsoft Graph activity followed?

Use this query to investigate service-principal sign-ins and Microsoft Graph API activity associated with the identified ConnectWise applications and suspicious IP addresses, starting October 13, 2026.

The query combines both event types into a chronological view, helping defenders identify suspicious authentication and related Graph operations, including potential credential changes, privilege modifications, and unusual resource access. Remove the IP filter for broader hunting. Results require further correlation, as authentication alone does not establish compromise. Run query

let Apps = dynamic([
    "506e08f9-6e9a-4da4-942b-a099a03644ae",
    "24205314-2b89-4e07-a2a4-579e898a5640",
    "0de14c77-b9b0-4137-9360-34878a929827"   
]);
let IPs = dynamic(["144.172.64.60", "144.172.64.99", "144.172.64.128"]);
union kind=outer
    (EntraIdSpnSignInEvents
     | extend EventType = "Sign-in"),
    (GraphAPIAuditEvents
     | extend EventType = "Graph API", IPAddress = IpAddress)
| where Timestamp >= datetime(2026-10-13)
| where ApplicationId in (Apps)
| where IPAddress in (IPs) // Comment out for broader hunting
| project-away IpAddress
| order by Timestamp asc

3 – Bulk Key Vault secret reads by a single identity across multiple vaults

Run query

CloudKeyVaultEvents
| where Timestamp > ago(14d)
| where ActivityName == "SecretGet" and HttpStatusCode == 200
| summarize SecretGets = count(), VaultCount = dcount(AzureResourceId),
   IPs = make_set(IPAddress, 20)
   by bin(Timestamp, 15m), AccountObjectId, AccountTenantId
| where SecretGets >= 20 and VaultCount >= 3

4 – Key Vault, Storage and ARM activity from the IOC IPs

Run query

let IPs = dynamic(["144.172.64.128","144.172.64.60","144.172.64.99"]);
union
 (CloudKeyVaultEvents | where IPAddress in (IPs) | extend Source = "KeyVault", Operation = ActivityName),
 (CloudStorageAggregatedEvents | where IPAddress in (IPs) | extend Source = "Storage", Operation = tostring(OperationNamesList)),
 (CloudAuditEvents | where IPAddress in (IPs) | extend Source = "ARM", Operation = OperationName)
| where Timestamp > ago(30d)
| project Timestamp, Source, IPAddress, Operation, AzureResourceId
| order by Timestamp asc

Sentinel hunting queries

1-SigninLogs → Did a service principal authenticate, from where, and against what resource?

Use SigninLogs to determine whether an affected application identity actually authenticated. This moves the investigation from potential exposure to observed authentication, allowing defenders to examine the application and service principal alongside the source IP address, geographic information, authentication result, and resource being accessed. Investigators should evaluate these sign-ins against the application’s expected activity and look for unusual source infrastructure, resources, or authentication patterns before pivoting into subsequent Graph activity. Run query

let IPIndicators = dynamic([
    "144.172.64.60",  // IP Indicator
    "144.172.64.99",  // IP Indicator
    "144.172.64.128"  // IP Indicator
]);
let ConnectWiseApps = dynamic([
    "506e08f9-6e9a-4da4-942b-a099a03644ae", // ConnectWise Cloud Backup
    "24205314-2b89-4e07-a2a4-579e898a5640", // ConnectWise - MS Teams
    "0de14c77-b9b0-4137-9360-34878a929827" // ConnectWise - MS Graph 
]);
SigninLogs
| where TimeGenerated >= datetime(2026-10-13)
| where AppId in (ConnectWiseApps)
//| where IPAddress in (IPIndicators) // Comment out to review all activity from these applications
| project
    TimeGenerated,
    AppId,
    ServicePrincipalName,
    ResourceDisplayName,
    ResourceId,
    IPAddress,
    CorrelationId
| order by TimeGenerated asc

2 – MicrosoftGraphActivityLogs → What observable Graph API activity followed?

Use MicrosoftGraphActivityLogs to investigate observable Microsoft Graph activity following authentication. This query allows defenders to pivot on the affected ApplicationId or tenant-specific ServicePrincipalId and examine details such as the request URI, HTTP method, response status, scopes, source IP address, and target workload. This helps establish what Graph activity occurred after the service principal authenticated and provides additional context around the resources the application attempted to access.

Importantly, the absence of results should not be treated as confirmation that no access occurred; Graph API telemetry should be considered alongside the application permissions and service principal authentication activity identified in the previous hunting steps. Run query

let IPIndicators = dynamic([
    "144.172.64.60",  // IP Indicator
    "144.172.64.99",  // IP Indicator
    "144.172.64.128"  // IP Indicator
]);
let ConnectWiseApps = dynamic([
    "506e08f9-6e9a-4da4-942b-a099a03644ae", // ConnectWise Cloud Backup
    "24205314-2b89-4e07-a2a4-579e898a5640", // ConnectWise - MS Teams
    "0de14c77-b9b0-4137-9360-34878a929827" // ConnectWise - MS Graph 
]);
MicrosoftGraphActivityLogs 
| where TimeGenerated >= datetime(2026-10-13)
| where AppId in (ConnectWiseApps)
| where IPAddress in (IPIndicators) // Comment out to review all Graph activity from these applications
| project
    TimeGenerated,
    AppId,
    IPAddress,
    RequestMethod,
    RequestUri,
    ResponseStatusCode,
    Scopes,
    ClientRequestId,
    OperationId,
    RequestId
| order by TimeGenerated asc

MITRE ATT&CK Techniques observed

This threat has exhibited use of the following attack techniques. For standard industry documentation about these techniques, refer to the MITRE ATT&CK framework.

Reconnaissance

• T1595 Active Scanning | Scanning cloud resources to identify targets for compromise.

Initial Access

• T1078.004 Valid Accounts: Cloud Accounts | Using compromised cloud accounts to obtain access.
• T1195 Supply Chain Compromise | Compromising a multi-tenant OAuth application to access downstream customer environments.
• T1199 Trusted Relationship | Access to customer environments through a compromised, trusted third-party application

Privilege Escalation

• T1098.003 Account Manipulation: Additional Cloud Roles | Addition of privileged cloud role assignments, including self assignment of global admin from one of the ConnectWise Management Application.
• T1098.001 Account Manipulation: Additional Cloud Credentials | Addition or modification of application credentials.

Credential Access

• T1528 Steal Application Access Token | Compromising OAuth application authentication material.
• T1555.006 Cloud Secrets Management Stores | Bulk secret reads from Azure Key Vaults

Discovery

• T1087.004 Account Discovery: Cloud Account | Identifying privileged cloud accounts for further targeting.
• T1580 Cloud Infrastructure Discovery | Mapping cloud infrastructure for further targeting.

References

• ConnectWise (SkyKick) Vendor Credentials Abused in M365 Attacks. Huntress
• ConnectWise Trust Center – Cloud Backup advisories. ConnectWise (accessed 2026-10-19).
• ConnectWise Cloud Backup technical investigation and remediation guide. ConnectWise (accessed 2026-10-19).

Learn more

• For guidance on investigating and responding to compromised applications, see Compromised and malicious applications investigation.
• Learn how to investigate app-only authentication with Service principal sign-in logs.
• Review application credentials, permissions, role assignments, and configuration changes in Microsoft Entra security operations for applications.
• Learn how to proactively investigate threats with Advanced hunting in Microsoft Defender XDR.
• Explore the OAuthAppInfo and GraphAPIAuditEvents advanced hunting tables.
• Learn how to find Microsoft 365 activity with Search the audit log.
• Review guidance for reducing excessive application permissions in Secure applications using least privilege.
• For the latest security research from the Microsoft Threat Intelligence community, check out the Microsoft Threat Intelligence Blog.
• To get notified about new publications and to join discussions on social media, follow us on LinkedIn, X (formerly Twitter), and Bluesky.
• To hear stories and insights from the Microsoft Threat Intelligence community about the ever-evolving threat landscape, listen to the Microsoft Threat Intelligence podcast.

The post AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise appeared first on Microsoft Security Blog.

Agents are rapidly becoming part of how work gets done. They can support a range of activities from writing code and researching products to analyzing revenue, processing expenses, and acting across the systems that run the business. To be most useful, they often need access to company data, applications, tools, and other agents. That access is what makes agents more capable. With that capability can come insider risk.

This does not mean agents are inherently malicious. It means they are capable tools that can be granted access to important systems by users. They can make mistakes, be manipulated, or be compromised. Their models can produce variable outputs, their context can change, and their plans can evolve as they work. An agent that operates as intended when a task begins may respond differently as context or inputs change. Depending on how they are designed and deployed, agents can operate continuously, at machine speed and scale.

Satya Nadella, Chairman and Chief Executive Officer of Microsoft, recently described the trust architecture this new era requires: we must separate the supply of intelligence from the authority over it. We should not ask a model to be its own security authority. The controls that determine what it can access, what actions it can take, and whether it should continue operating must sit outside the model, where the model cannot bypass or tamper with them.

Project Perception—Securing your agentic insider

We introduced Project Perception and the Microsoft cyber stack to help organizations defend with agents. Today, we are extending Perception to secure the agentic insider. Guardian Agents help security teams operate continuously, at machine speed and scale, while sensors, context, and actuators across the cyber stack help understand changing agent risk and continuously adapt protection in response. Project Perception and Microsoft Agent 365 share signals, context, and a unified registry giving IT and security teams a consistent view of agents and risk, with the controls each needs.

This builds on a simple security principle: surround non-deterministic intelligence with deterministic controls, independent evidence, and containment boundaries. Assume an agent can fail or be compromised. Observe it continuously. Minimize and tailor its authority before it acts. And continuously adjust that authority as risk changes.

This requires protection across identity, permissions, data, endpoints, network, cloud, and runtime. The system has four continuous responsibilities:

• Know which agents exist, who owns them, why they exist, and what they are doing.
• Contain what each agent can access and do before it acts.
• Protect the data, tools, runtime, resources, and systems in the agent’s path.
• Adapt protection continuously as behavior, context, and risk change.

These are not four products or sequential phases. They operate together. Evidence changes risk. Risk changes policy. Policy changes what the agent is allowed to do. Consider an expense-approval agent. It may need to read expense submissions, check policy, and approve routine reimbursements, but it should not inherit every permission of the employee who sponsors it. Its identity, runtime, access paths, and task-scoped authority should be explicit from the start. If it later encounters a prompt injection or attempts an action outside that task, the system should respond proportionately: preserve safe reads while blocking writes, narrow the resources it can reach, require provable human approval for a consequential step, or isolate and stop the agent when the risk warrants it. This is what continuous containment looks like in practice.

Know every agent and give it an accountable identity

You cannot secure what you cannot see. Agents can be created through Microsoft platforms, integrated through partner platforms and software development kits (SDKs), or introduced locally without a central deployment process. Microsoft Security uses native integrations, connectors, and sensors across the cyber stack to discover more of the agent estate.

We are expanding discovery to 35 additional types of local agents, bringing the total to 60 supported local agents, including Claude Code, OpenAI Codex, and GitHub Copilot. Organizations can bring these agents into a unified registry, regardless of how they were discovered, to understand ownership, activity, and risk.

Discovery tells you that an agent exists. Identity gives security and governance systems a durable anchor for controlling it.

Every agent should have a first-class identity distinct from the person using or sponsoring it. We are extending Microsoft Entra Agent ID with identity types for additional agent patterns, including local and previously unknown agents discovered through security sensors. This means an agent can become a policy target even when it was not originally integrated with Microsoft Security. Agent identity anchors attribution, ownership, lifecycle, risk, permissions, authorization, and policy, so the enterprise can distinguish what a person did from what an agent did on that person’s behalf.

Contain what the agent can do before it acts

Being able to find and identify an agent is necessary, but it is just the start. We need to understand the potential impact the agent can have in the enterprise environment. The next step is to bound the agent across its runtime, its paths to tools and resources, and the authority it can exercise.

Contain the runtime

The model and harness should execute inside a controlled environment rather than inherit fully-privileged access to the device. Sandboxes such as Microsoft Execution Container (MXC) provide policy-controlled access to resources the agent needs to complete its task. They also create an enforcement boundary where security controls and observability can operate outside the agent’s reach. If the agent or harness is compromised, it should not be able to disable the controls watching it.

Control the action space

An agent should not be free to call any tool, API, Model Context Protocol (MCP) server, resource, or other agent it can find. Local and network and data controls must mediate those paths, observe the calls, and enforce enterprise policy. Controlling the destination alone is not enough. Security must also understand the content flowing into and out of the agent, including its sensitivity, provenance, and whether untrusted data may be influencing the agent’s behavior.

Start with no permissions, or least privilege

We have spent decades applying Zero Trust and least privilege to people. Agents make those principles even more important. An agent should not automatically inherit all the standing access of its user. It should begin with no authority and receive only what it needs for the approved task, for only as long as the task requires.

To enable this, we are introducing Dynamic Permissions. Organization-wide policy sets the outer boundary, including preventing an agent from obtaining authority its user does not have. Permissions can then be scoped to the task, made time-bound, and contracted or removed as the task changes or completes. Higher-impact actions can require verified human approval.

Adapt protection as agent risk changes

Least privilege reduces exposure, but it does not eliminate risk. An expense-approval agent may begin a task with the right access and then encounter a prompt injection in an external document. Its configuration has not changed, but its behavior and risk have.

For human users, a delay between detection and response may be manageable. For an agent operating at machine speed, it can be the difference between containment and impact. The security system must be able to understand what changed and act before a person can manually investigate every signal.

That is why we are introducing Dynamic Agent Risk and Dynamic Conditional Access for agents.

Dynamic Agent Risk

Dynamic Agent Risk combines signals from identity, data, endpoints, cloud, network, runtime, and the agent’s own activity. It considers how securely the agent is configured, what resources it can reach, the potential impact if something goes wrong, and whether its actions are drifting from the user’s or organization’s intent.

The result is not a one-time score assigned at onboarding. It is a continuously updated understanding of the agent’s current risk and whether its current authority still matches that risk. The question is not simply whether the agent should run or stop. It is what the agent should be allowed to read, change, invoke, or approve at that moment.

Dynamic Conditional Access

Dynamic Conditional Access extends Microsoft Entra Conditional Access for agents so a change in risk can drive coordinated action across multiple enforcement points. The response is graduated, not binary. Policy can reduce permissions, preserve reads while blocking writes, narrow the data and resources the agent can reach, restrict network paths, shorten or revoke time-bound access, or require provable human approval before a consequential action proceeds. At the highest levels of risk, it can isolate the runtime or stop execution entirely.

Return to the expense agent. It begins with narrowly scoped permission to review and approve expenses within policy. When it encounters untrusted content and attempts an unrelated action, new evidence raises its Dynamic Agent Risk. Policy responds proportionately: it can preserve read access while removing the ability to approve or write, block access to sensitive resources, and require cryptographically provable human approval for a high-impact step. If the risk continues to rise, the system can isolate the runtime or stop execution. The employee remains productive. The agent is constrained without treating every risk change as a kill-switch event.

Secure agentic systems at scale with Guardian Agents

As organizations deploy thousands or hundreds of thousands of agents, manually triaging every alert and coordinating every response will not scale. Securing agents must itself become agentic, while remaining inside boundaries people set.

That is why we are introducing new Playbooks and a team of Guardian Agents in Project Perception. People define policy, approval boundaries, and delegation rules. Playbooks define the desired security outcome and coordinate protective actions. Guardian Agents carry out the work at machine speed, adapting as agent activity, risk, security context, and organizational policy change.

Guardian Agents can help teams understand why risk changed, investigate the relevant evidence, recommend or apply containment, and use what they learn to harden policy. This is not about outsourcing judgment. People retain control over objectives and consequential decisions. The system carries the continuous work that does not scale through manual operations.

Agents will only be adopted at scale if they can be trusted at scale. At Microsoft, we are already operating in a world of hundreds of thousands of agents. As Customer Zero, we have learned that securing agents requires continuous visibility into agent activity, proactive protection of access and data, and adaptive protection that keeps pace with evolving threats. Those lessons directly inform the innovations we are delivering to help every organization adopt agents with confidence.”

—Terrell Cox, Corporate Vice President (CVP) and Deputy Chief Information Security Officer (CISO), Customer Security, Microsoft

Extending containment to the model

As model transparency increases, it creates a new source of security evidence. Chain-of-thought and other model-level signals can help identify intent drift, understand whether untrusted information is influencing reasoning, and detect when an agent’s planned actions are diverging from the task it was given.

We are bringing this deeper evidence into Dynamic Agent Risk, together with signals from across the cyber stack. This builds on Satya Nadella’s principle: increase transparency into the model while keeping security decisions and enforcement independent of it.

Greater model transparency gives security systems deeper evidence to detect suspicious behavior and inform Dynamic Agent Risk. But the core security principle does not change: the model cannot be its own security boundary. Independent systems must observe every layer, including the model itself, and ultimately determine and enforce what the system is allowed to do.

These are not separate pillars. They make the same containment architecture stronger: better evidence, stronger assurance, more accurate risk, and more effective protection.

Confidence to move faster

Organizations are adopting agents to serve customers better, solve problems faster, and give people more time for the work that matters most. The goal is not to make every model perfectly predictable before it can be useful. The goal is to build systems that remain secure even when the model is not.

That requires a clear division of responsibility. Models supply intelligence. People and organizations retain authority. Independent controls establish identity, limit privilege, protect data, preserve evidence, and contain unsafe action. Security continuously reassesses risk and adapts protection as the agent works.

That is what we are working toward: an agentic world where organizations can put intelligence to work with confidence because security can observe it, govern it, and contain it at the same speed.

Learn more here and join us at Microsoft Ignite.

To learn more about Microsoft Security solutions, visit our website. Bookmark the Security blog to keep up with our expert coverage on security matters. Also, follow us on LinkedIn (Microsoft Security) and X (@MSFTSecurity) for the latest news and updates on cybersecurity.

The post Agents are your newest insiders. Security must move at their speed.  appeared first on Microsoft Security Blog.

The growing AI landscape is giving enterprises more choice in how they build, customize, and deploy increasingly capable models. Open-weight models are an important part of that landscape. They give enterprises access to model weights and greater flexibility to inspect, evaluate, modify, and run models within their own environments.

While that flexibility can be valuable, it also affects how enterprises evaluate models, where they implement safeguards, and who is responsible for maintaining those safeguards over time.

As enterprises choose which models to deploy, the relevant question is broader than whether a model is open-weight or closed-weight. Enterprises need to understand what they are using the model for, what could go wrong in that context, and what evidence and controls they need to manage risks.

Today, Microsoft and NVIDIA are publishing Risk Management for Open-Weight Models in Enterprise, the first in a planned series of joint resources intended to help enterprises address those questions. The paper introduces a shared risk taxonomy and an approach to managing model risks grounded in established standards and frameworks, including International Organization for Standardization (ISO) standards and the National Institute of Standards and Technology (NIST) AI Risk Management Framework.

The paper reflects a core idea: weight openness alone does not determine an AI system’s risk. Risk depends on the model, use case, deployment, and safeguards.

Understanding risk in the deployment context

Open- and closed-weight models can share many of the same underlying risks, but how those risks materialize depends on the deployment. A model used to help an employee draft a document presents different considerations than the same model used in a customer-facing workflow, just as a standalone inference endpoint differs from a system connected to external data or tools.

The architecture matters too. A risk may originate in the model, while the most effective control may sit in the inference environment, application, or surrounding operational processes. The joint guidance looks beyond the model label to factors including provenance, intended use, users, deployment conditions, and surrounding controls. This first paper focuses primarily on model-layer risks. Evaluating a model alone does not establish the safety of an agent or application: teams must also assess its permissions, data access, tools, runtime, and operational controls.

Establishing a common language

Managing these risks consistently also requires a common vocabulary across engineering, security, legal, safety, and business teams.

The initial Microsoft and NVIDIA taxonomy organizes risks into five broad categories: Cybersecurity, Data Protection, Content, Human-AI Interaction, and Bias. It is not a policy or deployment checklist, but instead it should be used by enterprises to adapt it to their own systems, use cases, and risk criteria, then connect relevant risks to evaluations, mitigations, and accountable owners.

Using evaluations as risk evidence

Evaluations help turn the taxonomy into something enterprises can use. The paper maps evaluations to a selection of risks across the five categories, giving teams a way to start with a risk that matters for their deployment and identify evaluations that can produce relevant evidence about it.

Microsoft Foundry helps enterprises assess quality, groundedness, and safety with built-in and custom evaluations and automated red teaming. Its configurable content-safety controls help detect and mitigate harmful inputs and outputs and prompt attacks. Teams should validate control availability for their model and deployment, including preview status, and combine these capabilities with application-specific testing, monitoring, and human oversight.

Evaluations still need to be interpreted in the context of a particular deployment and the goal is to make it easier for teams to connect the risks they identify with useful evidence, while building a broader set of evaluation resources that can evolve alongside models, deployment patterns, and emerging risks. Microsoft and NVIDIA plan to use the joint repository to share additional evaluations and point to existing resources that can help enterprises assess specific risks.

Accounting for changes in control and responsibility

Open-weight access gives enterprises greater control over how a model is evaluated, modified, and deployed. Model access and deployment are separate choices: open-weight models can also be used through managed platforms. Responsibility for safeguards depends on the service and deployment configuration; enterprises that operate their own model deployments may take on controls otherwise provided by a managed service.

Those safeguards can operate at different layers, and (depending on the risk) enterprises might use model changes, access controls, input and output filtering, monitoring, or human review. The important step is connecting each identified risk to controls that can effectively address it within the deployed system. Those decisions need to be revisited as models and deployments change.

Building on this work

Risk Management for Open-Weight Models in Enterprise focuses primarily on the model layer. Future work from Microsoft and NVIDIA is intended to extend this approach across the runtime, harness, and application layers, with additional technical guidance on evaluation and mitigation.

We also expect the resources supporting this work to evolve as new models, evaluations, and deployment patterns emerge. The joint repository will provide a place to share additional evaluation resources, including contributions from the broader community.

For enterprises deploying open-weight models today, the paper provides guidance they can begin applying to their own use cases. We encourage developers, security practitioners, researchers, and others working in this space to read the paper, explore the evaluations, and contribute evaluations that can help strengthen this shared body of work.

Read the joint Microsoft-NVIDIA paper, Risk Management for Open-Weight Models in Enterprise.

Explore the joint repository and technical resources.

Risk Management for Open-Weight Models in Enterprise was developed jointly by Microsoft and NVIDIA, with Andrew Berkley and April Chen of Microsoft and Michael Boone, Eileen Long, and Arun Rao of NVIDIA leading the development and editing of the paper.

Explore Microsoft Foundry evaluations and content-safety controls for implementation guidance.

To learn more about Microsoft Security solutions, visit our website. Bookmark the Security blog to keep up with our expert coverage on security matters. Also, follow us on LinkedIn (Microsoft Security) and X (@MSFTSecurity) for the latest news and updates on cybersecurity.

The post Managing open-weight model risk in enterprise deployments appeared first on Microsoft Security Blog.

Over the past year, Microsoft uncovered a Chinese-language cybercriminal ecosystem, an interconnected network of phishing kit operators, developers, and an illicit marketplace selling tools and infrastructure for large-scale phishing, enabling cybercriminal operations worldwide.

Threat actors purchased and used these tools and infrastructure, including phishing kits and access to compromised or fraudulent Microsoft or Google cloud environments and related credentials, to target organizations across the healthcare, financial services, wholesale distribution, and information technology (IT) sectors in Japan, the United States, Australia, and India. In collaboration with law enforcement agencies and other partners worldwide, Microsoft’s Digital Crimes Unit (DCU) facilitated a disruption of the ecosystem’s primary developers and infrastructure.

Microsoft’s research into one phishing kit unraveled a larger ecosystem, including the phishing kit developers, an infrastructure dealer, and the threat actors who buy from both. In this blog, we examine several phishing kits marketed and sold through YAM Store, an underground marketplace tracked by Microsoft as Storm-3119. We share our analysis of the technical aspects of this Chinese-language phishing kit ecosystem, the shared infrastructure, provisioning methods, and downstream activity by threat actors using these tools.

Uncovering the shadow world of a phishing kit ecosystem

Microsoft, alongside industry partner Resecurity, identified three phishing kits sold through the YAM Store that make up this Chinese-language cybercriminal network: CoGUI (also known as NX or FA), Haozi (also known as Phoenix), and Iron Man System. Microsoft tracks the cluster of activity and the developers behind all three phishing kits as Storm-2990. Since January 2026, the threat actors behind these kits have been responsible for facilitating the sending of over one billion phishing emails and SMS messages to over 75 thousand organizations worldwide.

Microsoft assesses with high confidence that these phishing kits were developed by Chinese-speaking threat actors, and, despite their distinct branding and separate Telegram-based sales channels, exhibit highly similar capabilities and functionality. The phishing kits provide phishing templates impersonating legitimate brands and organizations (including government, financial, and logistics) worldwide, and enable criminal actors to facilitate high-volume phishing and credential stealing operations. The similarity in their infrastructure and overlap in activity suggest either shared development origins, code reuse, or close collaboration among their developers.

These phishing kits have been advertised through YAM Store, first CoGUI and Haozi, followed by Iron Man System starting in June 2026. YAM Store and affiliated phishing kit companies accepted payments through cryptocurrency, including USDT and Ethereum, adding another layer to conceal illicit activity. The scale and proliferation of this ecosystem demonstrate how readily available criminal infrastructure can lower the barrier to entry, enabling less skilled cybercriminals to target organizations at scale and contribute to fraud losses.

By mapping the relationship between YAM Store and these phishing kits, we provide insight into the commercial infrastructure that supports modern phishing operations.

Storm-3119: YAM Store, the power broker

YAM Store is an initial access broker (IAB) that sells infrastructure using compromised or fraudulently obtained environments in providers like Microsoft and Google. The marketplace offers a variety of infrastructure based on age, subscription type, and environment types such as education (.edu) or non-profit (.org).

The YAM Store marketplace provides cybercriminals with a platform to purchase fraudulently acquired Microsoft and Google domains associated with legitimate, expired, or previously used organization numbers, allowing actors to use seemingly legitimate infrastructure and obtain workspaces at relatively low cost. Storm-3119 uses automation to identify expired environment or organization names and leverages legitimate customer information associated with them to help circumvent existing detections.

Anatomy of YAM Store operations: How Storm-3119 acquires, controls, and monetizes infrastructure

During this investigation, Microsoft uncovered a five-stage attack pipeline often used in Storm-3119’s operations. YAM Store functions as an access broker, an intermediary that hijacks Microsoft 365 and Azure for Education tenants and resells the stolen accounts, licenses, and storage as a service to downstream operations, including the phishing kits tracked as Storm-2990. The pipeline below describes the supply side of that business: how the actor acquires, controls, and monetizes compromised environments.

Stage 1 – Reconnaissance: Storm-3119 conducts automated reconnaissance of target environments. The actor uses Flask, a lightweight Python web framework, to covertly identify and profile cloud-enabled organizations. The tooling is designed to discover Microsoft 365 tenants, assess enrollment opportunities, gather licensing and Azure cloud information, and perform these activities anonymously through Tor infrastructure.

Stage 2 – Environment takeover and app-only admin panels: Once a viable environment is identified, the threat actor centralizes associated information in an actor-controlled platform used to administer compromised Microsoft 365 environments. This platform is a separate set of administrative panels (implemented in PHP, Java/Spring, and Cloudflare Workers) that lets the operator maintain, monetize, and expand access using stolen application credentials. Leveraging the compromised environment’s client secret through an application-only authentication flow, the panel performs full directory control: it resets passwords silently (without forcing a change at next sign-in), assigns education licenses, and grants privileged attributes that the actor can later monetize. Because this access is application-level rather than tied to a single user session, the resulting persistence is more durable, more scalable, and harder to detect.

Stage 3 – Storage and resource abuse: With the stolen privileged credentials, the threat actor signs in to a compromised Global Administrator account and self-provisions the PnP Management Shell service principal, which represents the actor’s application inside the compromised tenant. Through this service principal, the threat actor grants themselves broad Microsoft Graph permissions, including read and write access to directory data and AppCatalog.ReadWrite.All, along with full control over the SharePoint API. The actor then abuses the compromised tenant’s SharePoint and OneDrive as bulk hosting and converts the hijacked storage into a sellable product.

Stage 4 – Phishing and DNS infrastructure support: The threat actor maintains dedicated infrastructure to support and manage phishing operations on behalf of their customers. A custom DNS management platform enables centralized control of phishing domains, SSL certificates, and domain-routing configurations through Cloudflare.

Stage 5 – Storefront and monetization: Storm-3119 resells the hijacked accounts, licenses, and storage through an automated storefront, operating as the access broker that supplies these resources to downstream operations. Payment is accepted through Alipay, WeChat Pay, and USDT (TRC20), and the storefront supports a sub-reseller tier. Telegram is used for order notifications and buyer contact rather than as the primary sales channel.

In addition to targeting existing tenants, Storm-3119 also separately creates fraudulent organizations and supporting web presence to establish tenants on Microsoft’s platform, likely using AI-generated content and forged documentation to make the fabricated entities appear legitimate. In one case, the actor created an environment, likely using AI, named “Jewish Choice” under the guise of a nonprofit organization and associated it with jewishchoice[.]org, a domain registered only within the past several months that did not appear to represent a legitimate organization. This activity indicates that Storm-3119 likely established the website and associated domain to lend credibility to the fabricated nonprofit identity and facilitate tenant creation.

Storm-2990: The Iron Man System phishing kit

YAM Store sells its illicitly acquired infrastructure, as well as the prominent phishing kits tracked as Storm-2990, to other threat actors. The platforms are commercially licensed phishing-as-a-service (PaaS) products. They are designed to impersonate organizations across several industry verticals to steal information, including sign-in credentials (usernames, passwords) and financial data (credit card information), used to gain access to network environments.

Iron Man System’s customers purchase a subscription that provides access to the kit’s license server, with the domain hard-coded into the kit. This restriction adds a layer of obfuscation by limiting site access to paying customers. The kits allow attackers to steal money by abusing the stolen credit card information.

The Iron Man System phishing kit contains over 2,335 templates tailored to over 139 countries including Australia, Japan, India, and the United States, with the United States accounting for the most phishing templates. The phishing lures used by Iron Man System mimic various trusted brands and organizations, such as local government including authorities associated with traffic and parking fees, law enforcement, healthcare, finance, transportation, and retail. These lures are designed to trick targets into providing sensitive information. Most phishing lures used in the United States are transportation or government-related.

The phishing kits are capable of the following advanced evasion techniques to varying degrees:

• Geofencing, a technique used to restrict access to configured countries only, so only users from specific countries can access the phishing content from their web browser
• Fake CAPTCHA gates, used to prevent or block automated browsing from threat research tools by making the user click a box or execute an action
• IP address block list and allow list, used to block or allow IP addresses to prevent threat researcher interaction
• Anti-detection configuration related tosettings to evade security researcher detection
• Telegram bot configuration,including bot tokens and chat IDs for real-time phishing that enables credential exfiltration. It allows the kit operator to receive automatic notifications on their Telegram account when a target takes action on a phishing lure, including entering credentials, credit card data submission, and verification codes.

After the target provides the sensitive information (sign-in details, credit card info), the phishing kit collects it using a WebSocket, offering customers real-time victim interaction through Telegram notifications during credential harvesting. This confirms that the kit supports live, interactive phishing where the operator monitors each victim’s progress and manually controls what the victim sees next, a hallmark of sophisticated PaaS platforms. The attacker then harvests the information to use in follow-on fraudulent financial activities.

The Iron Man System phishing kit panel is predominately hosted on China-based hosting providers, with a smaller number of panels found on hosting providers in other countries. The fully featured, multi-tenant phishing kit uses the GoFrame application development framework, an open-source Go framework with roots in the Chinese developer ecosystem.

Downstream activity from Storm-2990 phishing kits

Microsoft observed downstream activity by threat actors using infrastructure affiliated with the Storm-2990 phishing kits. In analyzed cases, the actors used similar common phishing techniques, such as urgency in the message and attaching malicious links that redirect to actor-controlled infrastructure that allows these downstream threat actors to perform the following activities:

• Credential stealing: The actors use phishing pages that impersonate legitimate services to trick targets into providing credentials or other authentication information, enabling downstream actors to access compromised accounts.
• Device code phishing: In some cases, the actors abuse the device code authentication flow to trick victims into authenticating to legitimate applications, allowing downstream actors to obtain authentication tokens and access compromised accounts.
• Financial fraud: The end goal for the downstream actors is to use stolen payment data to make fraudulent purchases of high-end goods or fraudulent transactions on legitimate financial platforms.

The threat actors using the phishing kits conduct phishing campaigns at scale whether it is a missing toll payment or a traffic citation. The following screenshots show email or SMS phishing lures used in Storm-2990 phishing kits.

Mitigation and protection guidance

This interconnected network of Chinese-language phishing kit operators and developers operates as an enabling ecosystem, providing infrastructure and services that facilitate criminal activity rather than functioning as a standalone malware tool. Microsoft assesses that the ecosystem is not limited to the abuse of Microsoft services and extends to other providers.

While Microsoft notes that the identified organizations impacted by attacks stemming from Storm-3119 or Storm-2990 operations include financial, healthcare, and IT companies, the activity conducted by threat actors using these tools are common attacks that could affect any organization.

Most of the activity that stems from this phishing kit ecosystem consists of social engineering, phishing operations, and account and storage takeovers. Microsoft recommends the following to mitigate the impact of Storm-3119 and Storm-2990-related threats.

• In nearly all the cases observed by Microsoft, the phishing lures used by the Storm-2990 operators followed prescriptive scam red flags, including urgency, time-sensitive tasks such as clicking a provided link with official branding. Users should verify with the purported sender of the message or email, even when the task involves signing in to a known, secure account not provided by the link or calling the purported organization or company directly.
• Educate users about protecting personal and business information in social media, filtering unsolicited communication, identifying lures in spear-phishing emails and watering holes, and reporting reconnaissance attempts and other suspicious activities.
• Educate users about common phishing techniques. Sign-in prompts should clearly identify the application being authenticated to. As of 2021, Microsoft Azure interactions prompt the user to confirm (Cancel or Continue) that they are signing in to the app they expect, which is an option frequently missing from phishing sign-ins.
• Review our recommended settings for Exchange Online Protection and Microsoft Defender for Office 365 to ensure your organization has established essential defenses and knows how to monitor and respond to threat activity.
• Configure the Microsoft Defender for Office 365 Safe Links policy to apply to internal recipients.

Hardening credentials and cloud identities is also necessary to defend against phishing attacks, which seek to gain valid credentials and access tokens. As an initial step, use passwordless solutions like passkeys and implement multifactor authentication (MFA) throughout your environment:

• Invoke the principles of least privilege and Zero Trust to significantly slow down threat actors who might have been able to gain initial access and buy time for defenders to respond. Since Storm-3119 operators have been observed obtaining Global Administrator access, implementing these principles can help lock down Global Administrator roles and who has access to specific resources.
• Use the Microsoft Authenticator app for passkeys and MFA, and complement MFA with Conditional Access policies, where sign-in requests are evaluated using additional identity-driven signals.
• Conditional Access policies can also be scoped to strengthen privileged accounts with phishing-resistant MFA.
• Your passkey and MFA policy can be further secured by only allowing MFA and passkey registrations from trusted locations and devices.
• Enable network protection to prevent applications or users from accessing malicious domains and other malicious content on the internet.

Microsoft Defender detections

Microsoft Defender customers can refer to the list of applicable detections below. Microsoft Defender coordinates detection, prevention, investigation, and response across endpoints, identities, email, and apps to provide integrated protection against attacks like the threat discussed in this blog.

The coverage below applies to the downstream activity detected by Microsoft that was identified using Storm-2990 infrastructure.

Microsoft Security Copilot

Microsoft Security Copilot is embedded in Microsoft Defender and provides security teams with AI-powered capabilities to summarize incidents, analyze files and scripts, summarize identities, use guided responses, and generate device summaries, hunting queries, and incident reports.

Customers can also deploy AI agents, including the following Microsoft Security Copilot agents, to perform security tasks efficiently:

• Threat Intelligence Briefing agent
• Phishing Triage agent
• Threat Hunting agent
• Dynamic Threat Detection agent

Security Copilot is also available as a standalone experience where customers can perform specific security-related tasks, such as incident investigation, user analysis, and vulnerability impact assessment. In addition, Security Copilot offers developer scenarios that allow customers to build, test, publish, and integrate AI agents and plugins to meet unique security needs.

Threat intelligence reports

Microsoft Defender XDR customers can use the following threat analytics reports in the Defender portal (requires license for at least one Defender XDR product) to get the most up-to-date information about the activity and techniques discussed in this blog. These reports provide intelligence, protection information, and recommended actions to prevent, mitigate, or respond to associated threats found in customer environments.

• Threat Overview Profile: Evolving phishing threats

Microsoft Security Copilot customers can also use the Microsoft Security Copilot integration in Microsoft Defender Threat Intelligence, either in the Security Copilot standalone portal or in the embedded experience in the Microsoft Defender portal to get more information about the activity discussed in this blog.

Hunting queries

Microsoft Defender XDR

Microsoft Defender XDR customers can run the following advanced hunting queries to find generic activity that may have some overlap with Storm-2990-related activity. Note that these queries are not specific to Storm-2990:

Hunt for suspicious device code authentication

The following query identifies risky sign-ins using the device-code authentication flow that could indicate device-code phishing and subsequent unauthorized account access.

EntraIdSignInEvents
| where Timestamp > ago(30d)
| where AuthenticationProtocol =~ "deviceCode"
| where RiskLevelDuringSignIn in~ ("medium", "high")
or RiskLevelAggregated in~ ("medium", "high")
or RiskEventTypes_V2 != "[]"
| project Timestamp, AccountUpn, IPAddress, Country, Application, ResourceDisplayName,
AuthenticationProtocol, RiskLevelDuringSignIn, RiskLevelAggregated,
RiskEventTypes_V2, SessionId
| order by Timestamp desc

Hunt for potentially compromised accounts using multiple risk signals

The following query identifies accounts exhibiting risky authentication activity across multiple IP addresses or countries that may indicate credential compromise following phishing.

EntraIdSignInEvents
| where Timestamp > ago(30d)
| where RiskLevelDuringSignIn in~ ("medium", "high")
    or RiskLevelAggregated in~ ("medium", "high")
    or RiskEventTypes_V2 != "[]"
| summarize FirstSeen = min(Timestamp), LastSeen = max(Timestamp),
            IPAddresses = make_set(IPAddress, 20), Countries = make_set(Country, 20),
            Applications = make_set(Application, 20), RiskEvents = make_set(RiskEventTypes_V2, 20),
            SignInCount = count() by AccountUpn
| where array_length(IPAddresses) > 1 or array_length(Countries) > 1
| order by LastSeen desc

Hunt for phishing clicks followed by Defender alerts

The following query correlates email URL clicks with subsequent Defender alerts for the same user to identify potential phishing-related account compromise.

let Lookback = 30d;
let PhishClicks =
    UrlClickEvents
    | where Timestamp > ago(Lookback)
    | project ClickTime = Timestamp, AccountUpn, Url, ClickAction = ActionType;
let IdentityAlerts =
    AlertEvidence
    | where Timestamp > ago(Lookback)
    | where EntityType =~ "User"
    | project AlertTime = Timestamp, AccountUpn, AlertId, Title, Severity, DetectionSource;
PhishClicks
| join kind=inner IdentityAlerts on AccountUpn
| where AlertTime between (ClickTime .. ClickTime + 24h)
| project ClickTime, AlertTime, AccountUpn, Url, ClickAction, AlertId, Title, Severity, DetectionSource
| order by ClickTime desc

Hunt for phishing URL clicks followed by suspicious sign-ins

The following query identifies users who clicked URLs in email and subsequently generated risky sign-in activity, which may help identify credential compromise following phishing.

let Lookback = 30d;
let PhishClicks =
    UrlClickEvents
    | where Timestamp > ago(Lookback)
    | where ActionType in~ ("ClickAllowed", "ClickBlocked")
    | project ClickTime = Timestamp, AccountUpn, Url, Workload, ClickAction = ActionType;
let SuspiciousSignins =
    EntraIdSignInEvents
    | where Timestamp > ago(Lookback)
    | where RiskLevelDuringSignIn =~ "high"
        or RiskLevelAggregated =~ "high"
        or RiskEventTypes_V2 != "[]"
    | project SignInTime = Timestamp, AccountUpn, IPAddress, Country, Application,
              RiskLevelDuringSignIn, RiskLevelAggregated, RiskEventTypes_V2;
PhishClicks
| join kind=inner SuspiciousSignins on AccountUpn
| where SignInTime between (ClickTime .. ClickTime + 24h)
| project ClickTime, SignInTime, AccountUpn, Url, IPAddress, Country, Application,
          RiskLevelDuringSignIn, RiskEventTypes_V2
| order by ClickTime desc

Microsoft Sentinel

Microsoft Sentinel customers can use the TI Mapping analytics (a series of analytics all prefixed with ‘TI map’) to automatically match the malicious activity mentioned in this blog post with data in their workspace. If the TI Map analytics are not currently deployed, customers can install the Threat Intelligence solution from the Microsoft Sentinel Content Hub to have the analytics rule deployed in their Sentinel workspace.

• Possible phishing with clicked/selected link and network session
• Threat Intelligence URL match with UrlClickEvents
• Correlate unfamiliar sign-in properties and atypical travel alerts
• Possible AiTM phishing attempt against Microsoft Entra ID
• Business email compromise – Mailbox rule

References

• CoGUI Phish Kit Targets Japan with Millions of Messages | Proofpoint US
• IronToll: Global Government-Impersonation PhaaS | PhishEye

Learn more

For the latest security research from the Microsoft Threat Intelligence community, check out the Microsoft Threat Intelligence Blog.

To get notified about new publications and to join discussions on social media, follow us on LinkedIn, X (formerly Twitter), and Bluesky.

To hear stories and insights from the Microsoft Threat Intelligence community about the ever-evolving threat landscape, listen to the Microsoft Threat Intelligence podcast.

The post The hidden network: Inside the phishing infrastructure ecosystem powering global cybercrime appeared first on Microsoft Security Blog.

Post-quantum cryptography (PQC) planning often focuses on protecting encrypted data, but authentication systems must also evolve. Certificates, trust anchors, PKI services, applications, devices, and hardware security modules may all be affected by new algorithms and larger certificate chains.

Microsoft’s Post-Quantum Cryptography (PQC) Transport Layer Security (TLS) Pilot Program helps eligible certificate authorities evaluate interoperability and operational readiness in controlled environments. Organizations should begin now by inventorying certificate dependencies, assessing vendor roadmaps, identifying long-lived infrastructure, and creating non-production test environments. Early testing can surface compatibility and process gaps before post-quantum authentication is required at scale.

By the end of this article, security leaders, public key infrastructure (PKI) administrators, and architects will understand the key certificate dependencies, migration risks, and practical first steps for post-quantum authentication readiness.

Why post-quantum authentication presents a different challenge

Much of the current PQC readiness discussion focuses on confidentiality and the ‘harvest now, decrypt later’ risk, in which adversaries collect encrypted data today with the expectation that future quantum computers could decrypt it.

Authentication introduces a different set of challenges.

While confidentiality protections primarily focus on safeguarding encrypted communications, authentication depends on a broad ecosystem of technologies and trust relationships. Certificates and private keys must be issued, distributed, stored, validated, renewed, and managed across diverse environments and vendors. Changes to cryptographic algorithms can therefore have implications that extend beyond cryptographic correctness into operational processes, interoperability, and infrastructure readiness.

Organizations frequently understand where TLS protects their communications. Fewer organizations have a complete inventory of every system that issues, validates, stores, distributes, or depends on certificates.

As the industry moves toward post-quantum authentication, security teams will need greater visibility into those dependencies.

Why organizations should begin preparing now

The transition to post-quantum authentication is unlikely to be a single technology upgrade.

Enterprise environments often contain decades of accumulated infrastructure, including internally managed PKI deployments, embedded devices, operational technology, security appliances, custom applications, third-party services, and hardware-backed trust systems. Some of these technologies may have long deployment lifecycles, fixed cryptographic assumptions, or operational constraints that are not immediately visible until testing begins.

The challenge facing organizations is not that post-quantum standards are unavailable. Rather, it is understanding how proposed post-quantum certificate hierarchies interact with the systems and processes already deployed across their environments.

Testing helps answer questions such as:

• Can existing applications correctly process and validate post-quantum certificates?
• How do larger post-quantum certificates and certificate chains affect handshake size, performance, storage, transmission, and inspection limits?
• Do enterprise PKI workflows require changes to accommodate new algorithms?
• Are network monitoring, inspection, or certificate-management systems prepared for post-quantum authentication?
• Are hardware security modules and other cryptographic infrastructure components ready to support future certificate requirements?
• What hidden dependencies exist within supply chains, vendors, and third-party services?

In many cases, organizations may not yet know which of these questions are most relevant to their environments. Identifying those unknowns is one of the primary reasons ecosystem testing is important today.

Microsoft’s perspective on readiness

Post-quantum authentication readiness requires validating how certificate chains, platforms, and operational processes behave together.

The goal is not simply to determine whether a certificate can be issued or validated. The goal is to identify interoperability, compatibility, performance, and operational challenges before post-quantum authentication must be deployed at scale.

Organizations should view post-quantum readiness as a multi-year planning effort rather than a future migration project. Security leaders who begin inventorying dependencies, assessing vendor readiness, and testing interoperability now will be better positioned to make informed decisions as standards, platform support, and industry requirements continue to evolve.

Testing the future certificate ecosystem

To help the ecosystem gain practical experience with post-quantum authentication, Microsoft launched the PQC TLS Pilot Program on August 27, 2026.

The pilot enables approved certificate authorities in good standing with the Microsoft Trusted Root Program to evaluate PQC TLS roots and certificate issuance using quantum-resilient Module-Lattice-Based Digital Signal Algorithm, ML-DSA-87.

The program is intended to create a controlled environment where participants can evaluate interoperability, compatibility, performance, and operational considerations associated with post-quantum certificate hierarchies.

Certificates issued through the pilot are not publicly trusted and are intended solely for interoperability and ecosystem-readiness testing in closed environments, custom applications, and enterprise testbeds. They must not be used for production trust scenarios or public-facing websites. ML-DSA support remains limited and evolving during the pilot and should not be relied upon for production use.

According to the Microsoft Trusted Root Program Annoucement, the August 2026 release added seven pilot roots operated by ComSign, DigiCert, HARICA, IdenTrust Services, Sectigo, Shanghai Electronic Certification Authority, and SSL.com.

The pilot will continue with rolling admissions through the end of 2026, providing additional opportunities for eligible certificate authorities participating in the Microsoft Trusted Root Program to join. Organizations that participate in the Microsoft Trusted Root Program and are interested in evaluating post-quantum authentication readiness should review the PQC TLS Pilot Program requirements and application process through the https://aka.ms/rootcert. The pilot provides eligible certificate authorities with an opportunity to test interoperability, operational processes, and ecosystem readiness ahead of broader post-quantum authentication adoption.

The objective is not to validate a predetermined outcome. It is to help the certificate ecosystem discover where modernization, testing, and operational improvements may be required before post-quantum authentication can be adopted broadly.

Expanding end-to-end testing

On supported and appropriately configured Windows 11 systems, ML-DSA certificates can be evaluated in the pilot’s controlled, non-production testing scenarios. Support begins with the July 28, 2026 updates: KB5101681 (OS Build 28000.2608) for 26H1 and KB5101684 (OS Builds 26200.8973 and 26100.8973) for 25H2.

ML-DSA certificates can also be used with Secure Channel (Schannel) in supported scenarios. Confirm the applicable Windows and Schannel platform requirements before testing.

Together, Microsoft platform support and the PQC TLS Pilot Program provide a controlled way to identify application incompatibilities, certificate-size constraints, performance effects, and operational workflow gaps before they affect production environments.

What organizations can do today

Organizations do not need to wait for broad industry adoption to begin preparing for post-quantum authentication.

Security leaders, PKI administrators, and architects can begin by:

1. Inventory certificate-dependent systems. Identify applications, services, devices, appliances, and infrastructure that depend on certificates for authentication, trust establishment, or lifecycle management.

2. Map trust relationships.

Document both public and private PKI environments, including internal certificate hierarchies, trust anchors, device authentication systems, and externally managed certificates.

3. Assess vendor readiness.

Engage certificate providers, PKI vendors, hardware security module (HSM) providers, software vendors, and platform providers to understand their post-quantum roadmaps and testing capabilities.

4. Identify long-lived infrastructure.

Pay particular attention to systems with lengthy upgrade cycles, including embedded devices, appliances, operational technology environments, and security infrastructure.

5. Develop a safe testing strategy.

Establish non-production environments where post-quantum certificate hierarchies can be evaluated for interoperability, performance, and operational impacts.

6. Build a multi-year transition roadmap.

Treat post-quantum authentication readiness as a program rather than a project. Assign owners, sequence dependencies, and use test results to prioritize modernization work before future requirements emerge.

Organizations interested in testing should consult their certificate provider about whether it participates in the pilot and whether it offers testing opportunities appropriate for their environment.

If your certificate provider participates in the Microsoft Trusted Root Program but is not currently part of the pilot, consider encouraging them to participate through the program’s admissions process.

Looking ahead

Post-quantum authentication will require coordinated changes across certificate authorities, platforms, software, hardware, and enterprise environments. Organizations that begin testing trust infrastructure and certificate dependencies now will be better positioned to reduce future migration risk and strengthen readiness for a quantum-resistant future.

Learn more

• PQC TLS Pilot Program Requirements
• Microsoft Trusted Root Program Announcements
• New Windows Features to Secure Today’s Data in a Post-Quantum World | Microsoft Community Hub
• ASP.NET, Kestrel and Schannel GA with TLS 1.3 Post-Quantum Cryptography
• Quantum-safe security: Progress towards next-generation cryptography
• Microsoft’s quantum-resistant cryptography is here | Microsoft Community Hub
• Post-Quantum Cryptography APIs Now Generally Available on Microsoft Platforms
• For the latest security research from the Microsoft Threat Intelligence community, check out the Microsoft Threat Intelligence Blog.
• To get notified about new publications and to join discussions on social media, follow us on LinkedIn, X (formerly Twitter), and Bluesky.
• To hear stories and insights from the Microsoft Threat Intelligence community about the ever-evolving threat landscape, listen to the Microsoft Threat Intelligence podcast.

The post Post-quantum authentication: Why organizations should start testing certificate ecosystems now appeared first on Microsoft Security Blog.

The mission of Microsoft Security’s Frontier Offensive Research & Generative Exploitation (FORGE) Lab is to advance the frontier of autonomous security engineering. We’re building a team that enables AI-native vulnerability research at Microsoft, pushing the boundaries of finding and fixing zero-day vulnerabilities. Four principles guide that work: autonomy over labor, defense through offense, building ecosystems over individual examples, and understanding over findings. This post reflects those principles in practice across Windows, the Linux kernel, and widely used open-source projects.

From May 2026 through September 2026, FORGE helped discover Windows vulnerabilities assigned 140 common vulnerabilities and exposures (CVEs), including 52 addressed in September 2026’s security release alone. The work reaches well beyond Windows. FORGE members submitted 155 internally validated reports across 23 open-source projects, including the Linux kernel. And through Akrites, a Linux Foundation initiative that coordinates confidential remediation and disclosure for vulnerabilities in critical open-source software, one of our Linux reports became the first Akrites submission to result in a patch merged into the Linux kernel.

These results demonstrate that agentic discovery can operate at meaningful volume, but they also expose the next constraint. Finding a difficult vulnerability proves an agent’s frontier capability. Repeatedly converting such findings into security updates requires a different kind of system—one that can move credible work through validation, remediation, and release. Discovery creates security value only when validation and remediation can keep pace. For security leaders, the question is no longer only whether AI can find vulnerabilities, but whether an organization can validate and fix them as fast as they are found.

Three lessons follow, each marking a shift in how we approach vulnerability research:

1. From frontier capability to scale.
2. From token consumption to reasoning economics.
3. From isolated discovery to coordinated validation and remediation.

The scarce resource is not always model intelligence. It can be a working build and deployment, a reproducible trigger, or even an engineer’s time.

At FORGE Lab, we use the multi-model agentic scanning harness, codename MDASH, to organize this work. Our May 2026 introduction detailed its design, while the June 2026 update covered pipeline improvements and benchmark analysis. Here, we focus on what those experiences teach us about operating vulnerability research at scale—across Windows and open-source projects.

Shift 1. From frontier capability to scale

The frontier question is whether a system can find a bug that demands deep reasoning, as we explored in our early work on the CyberGym benchmark. The scale question is whether it can repeat that result across targets without rebuilding every environment, investigation, and review process from scratch. Better models still matter, especially for difficult or unfamiliar bug classes. But once discovery becomes repeatable, model capability is only one constraint on useful output.

Adding auditors can increase candidate volume without increasing the rate of validated findings or shipped fixes. If reports arrive faster than reviewers—in our case, Microsoft Security Response Center (MSRC)—can resolve them, the immediate result is a growing queue. A larger search budget can even reduce useful throughput when duplicate or poorly supported reports consume attention that stronger findings need.

The unit to optimize, therefore, is neither the scan nor the report. It is a reproducible finding that advances with enough evidence for the next stage to act. Reusable preparation, deduplication, and review capacity are therefore core components of the research system.

As agentic systems produce more candidates, the bottleneck shifts from discovery to determining which reports are real, reachable, and security-relevant. Project-specific automated provers—such as proof of vulnerability (PoV)/proof of concept (PoC) generators, harness builders, and trigger-input finders—turn plausible reports into reproducible evidence that engineers and maintainers can act on. By finding crashing inputs, confirming reachable execution paths, and producing regression-ready triggers within each project’s build and test environment, these tools can help reduce human triage, accelerate remediation, and make vulnerability research sustainable at scale. For example, by leveraging deterministic algorithms such as abstract syntax tree (AST), one of our internal projects reduced about 45% duplicate findings across multiple scans of the same code which reduced the load on the PoC generator and subsequently human triage effort.

More candidates do not create more throughput when the review queue grows faster than fixes ship.

Shift 2. From token consumption to reasoning economics

At scale, every repeated orientation, speculative branch, and redundant debate carries a cost. But minimizing tokens alone is the wrong objective. A short, ambiguous report may be cheap to generate yet expensive to investigate; a longer analysis that establishes the missing execution path may reduce total system cost.

The key question is where the next unit of reasoning will change a decision. If an index can identify callers, asking a frontier model to rediscover them wastes capacity. If the uncertainty is whether two lifetime conditions can coexist across callbacks, deeper reasoning may be warranted. If the question is whether an input triggers the failure, an executable check can provide evidence that more prose cannot.

MDASH combines frontier and distilled models, specialized auditors, and code-analysis tools, enabling work to be allocated by task. That flexibility does not guarantee optimal allocation. Routing routine work to cheaper models, reusing verified context, and escalating unresolved questions to stronger reasoning are hypotheses to test—not efficiency gains to assume.

A useful allocation policy starts by identifying what remains unknown: a caller, a build configuration, a reproducer, or a causal explanation. The next action should close that specific evidence gap. Repeating a review without adding evidence spends more tokens while preserving the same uncertainty. Early filtering can also discard real bugs, so any savings must be measured against coverage and missed findings.

Scan outcomes can also become training data. At scale, vulnerability scanning should become a training loop, not only a discovery pipeline. Each MDASH run produces signals that can improve future models: true-positive and false-positive verdicts, duplicate findings, failed reachability claims, reviewer feedback, verifier results, severity assessments, patch outcomes, and regression-test evidence. Capturing those signals with the code context and causal argument behind each candidate creates the dataset needed for reinforcement learning and fine-tuning specialized cyber models. The goal is for future models to learn not just what a vulnerability looks like, but which findings survive validation, which explanations help humans and provers act, and which patterns lead to useful remediation.

Spend reasoning to remove uncertainty, not simply to produce more analysis.

Shift 3. Validation and remediation as a continuous learning loop

Validation and remediation are not downstream cleanup; they are part of the discovery loop. Each candidate should move through automated verification, human review, patch development, and regression testing, with every stage returning evidence to the system. A reproducible trigger strengthens the report, guides the fix, and can seed a regression test. A failed validation is also useful when it records why: an unreachable path, missing precondition, incorrect build configuration, insufficient attacker control, duplicate report, or incomplete causal explanation.

That evidence is how the system improves. Project-specific provers, harnesses, and trigger generators accumulate operational knowledge about how each target builds, runs, fails, and accepts fixes. Remediation outcomes show the discovery pipeline which evidence changed a decision, which assumptions failed, and which bug patterns warrant more or less attention. The objective is to make every investigation leave reusable capacity behind: a validated trigger, clearer invariant, better harness, stronger regression test, or routing rule that prevents the same mistake.

Human attention remains essential, but it should concentrate where judgment has the highest leverage: assessing security impact, reviewing patches, and determining whether a change restores the component’s intended invariant. Automation can handle the repeatable work of establishing reachability, reproducing behavior, and preserving evidence. Over time, this loop converts individual findings into system knowledge—raising the quality of future discovery while shortening the path from credible report to shipped fix.

A reproducible defect is not yet a serviced fix. To shorten the path from discovery to remediation, MDASH needs to operate as part of the engineering loop: continuously scanning code, connecting findings to the builds and artifacts produced by continuous integration and continuous delivery (CI/CD), and feeding validation results back into development, as described in the Windows team’s blog post. Broad analysis can identify suspicious code paths, but component-specific proving needs the right binaries, symbols, configurations, harnesses, and runtime conditions to reproduce a candidate defect against the version that matters for customers and releases. Each handoff should be both owned and machine-usable: research supplies the candidate claim, causal path, and uncertainty; proving supplies execution evidence from the relevant build artifact; component teams assess the violated invariant, repair strategy, compatibility risk, and related variants; and servicing connects the approved change to release validation. A failed reproduction, rejected finding, incomplete patch, or regression result should feed back into MDASH as structured evidence, not simply move a ticket into another queue. The goal is not to remove human code review or release validation, but to automate the repeatable work around them—finding candidates, selecting artifacts, reproducing behavior, preserving evidence, suggesting related paths, and returning what was learned to the next scan.

The goal is to build an organization-wide system that automates repeatable work while preserving clear engineering accountability.

Transferring security research to help open source projects

In open source, all three constraints apply at once. Each project has its own setup costs, mix of analysis and execution work, and maintainer workflows and review capacity. Within Windows, vulnerability research connects to established component owners and servicing infrastructure. Across open-source projects, reusable methods can reduce repeated setup, but they cannot replace the technical and human context required to move a finding toward a fix.

Open-source vulnerability reporting snapshot

Over three months, FORGE members audited open-source projects spanning kernels, runtimes, networking libraries, container technologies, and media parsers. The team submitted 155 internally validated reports across 23 projects. At the time of writing, 93 reports across 14 projects or project families had documented maintainer acknowledgement or acceptance: apple/container, apple/containerization, curl, Escargot, FFmpeg, Hyperlight, Linux, llama.cpp, Node.js, PyRIT, Rust, SQLite, vLLM, etc.

Since the reports are at different stages of that process, only subsets are publicly disclosed today, such as curl’s CVE-2026-9545 and CVE-2026-13608, Node.js’s CVE-2026-56848, and Linux kernel CVE-2026-64563. Our public CVE index lists the CVEs currently cleared for disclosure, and you can find all our public channel Linux kernel reports through lore.kernel.org query.

Measured costs of automated validation

MDASH surfaced thousands of suspicious locations in the Linux kernel, and our validation agents produced supporting evidences for 627 findings. The figures below summarize automated validation costs for a subset of these efforts. Across 182 confirmed crash findings, generating a PoC averaged $3.61 in model cost and 21.5 minutes. For six selected cases, using automatic exploit generation (AEG) to test local privilege-escalation potential averaged $8.56 and 25.4 minutes.

These evaluations used GPT-5.5, without extensive task-specific tuning of either the model or the harness. These averages include all attempts associated with the successful cases shown, but exclude initial screening, candidates that failed or were filtered out, and human investigation and patch preparation. Within those limits, the results still strongly indicate that automated validation can produce useful results at practical cost and latency, even at Linux-kernel scale. Tighter co-design of models and agent harnesses could improve both validation yield and efficiency. If costs fell by another order of magnitude, for example, 10x or even 50x, the range of findings that could be economically validated would expand substantially, further reshaping vulnerability research.

FORGE OSS Bug Hunt Party: Human coordination at scale (Hackathon)

Looking back to those efforts, technical evidence was only part of the job. Getting it fixed also required understanding and respecting each project’s context and workflow: some required a patch while others did not; some preferred private disclosure while others used public channels; and maintainers sometimes differed on whether an issue crossed a security boundary. In practice, this meant adapting to each project’s submission requirements, working with maintainers to reproduce the issue, and, where applicable, refining and retesting a proposed fix. As AI-assisted discovery increases report volume, open-source security research can scale only if research teams absorb the resulting complexity rather than pass it on to maintainers.

Scaling this work also requires more developers who understand the full security workflow. At Microsoft’s annual Hackathon, we hosted the FORGE Open Source Software (OSS) Bug Hunt Party (Project Sunshine) as a forum where developers across Microsoft could exchange security research practices and gain hands-on experience with vulnerability investigation, validation, and responsible reporting. With MDASH and supporting guidance, 50 participants produced 39 reports across six projects, including Microsoft’s open-source Hyperlight project, Linux, vLLM, Gemini CLI, and llama.cpp. Those 39 reports are included in the 155-report snapshot above. Project Sunshine has continued beyond the Hackathon as a forum for open-source security collaboration within Microsoft.

Akrites: Coordinated effort to remediate vulnerabilities

Akrites is a Linux Foundation initiative that coordinates confidential remediation and disclosure for vulnerabilities in critical open-source software. FORGE has submitted nine Linux findings with internally assessed common vulnerability scoring system (CVSS) scores above 7.0. Together with submissions from other participants, including Google, these cases helped test and refine Akrites’ early Coordinated Vulnerability Disclosure (CVD) workflow. One of those reports was the first Akrites submission to result in a patch merged into the Linux kernel. FORGE will keep contributing to those responsible disclosure efforts.

The work also showed that a systematic approach is not just a larger scan. It is a repeatable path from a finding to an upstream fix and coordinated disclosure. Each case needs reproducible evidence, a clear causal explanation, a severity assessment, and an understanding of which downstream projects may be affected. The right owners and collaborators can then be brought in at the right time and on a need-to-know basis.

What’s next for FORGE Lab?

In a matter of months, FORGE Lab has helped discover Windows vulnerabilities and assigned 140 CVEs, including 52 in September 2026’s security release. Beyond Windows, FORGE submitted 155 internally validated reports across 23 open-source projects—93 of which had documented maintainer acknowledgement or acceptance at the time of writing—and one of our Linux reports became the first Akrites submission to result in a patch merged into the Linux kernel. Our Linux kernel results also indicate that automated validation can work at practical cost, with proofs of concept averaging $3.61 in model cost and 21.5 minutes per successful case. The lesson behind those numbers: once agents can find difficult bugs, progress depends on the system around them, how it scales, where it spends reasoning, and how quickly it turns discoveries into validated fixes.

These three lessons are becoming foundational guidelines for FORGE’s next phase of research:

1. Moving from frontier capability to scale means measuring the full pipeline, not just the number of findings: candidate arrivals, duplicates, rejected reports, validated defects, queue age, unresolved cases, and time through each stage.
2. Moving from token consumption to reasoning economics means evaluating model and execution cost alongside human review time, and asking whether each additional pass removes uncertainty, improves coverage, or produces independently validated findings.
3. Moving from isolated discovery to continuous validation and remediation means tracking reproduction success, patch rework, regression evidence, and elapsed time from validated defect to approved fix and release.

For security leaders evaluating AI-powered vulnerability discovery, these are useful measures too: count validated fixes, not just findings, and weigh model cost alongside the human review time it takes to act on them.

The next advance won’t come from a better model alone. It will come from research systems that combine frontier capability, systems engineering, and reasoning economics, focusing machine reasoning and human attention on turning credible findings into validated fixes.

Our team is growing. Check out our open positions at FORGE Lab.

Further reading:

• Defense at AI speed: Microsoft’s new multi-model agentic security system tops leading industry benchmark.
• Beyond the benchmark: Advancing security at AI speed.
• Evolving Windows vulnerability management to meet the speed of AI-powered discovery.
• From research to reality: An interview with Microsoft VP of Security Research Taesoo Kim.
• Microsoft Security FORGE Lab.
• Berkeley Vulnerability Initiative—Agentic Vulnerability Coverage Map.
• Akrites | Patch the Commons, Together.
• Introducing MAI-Cyber-1-Flash inside MDASH | Microsoft AI.

To learn more about Microsoft Security solutions, visit our website. Bookmark the Security blog to keep up with our expert coverage on security matters. Also, follow us on LinkedIn (Microsoft Security) and X (@MSFTSecurity) for the latest news and updates on cybersecurity.

The post 3 lessons from frontier AI vulnerability research appeared first on Microsoft Security Blog.

Most of what has been written about AI and vulnerability management focuses on speed: how much faster frontier AI models can scan code, find weaknesses, design patches, and build exploits than any human team. That part is true, and it matters to how we remediate vulnerabilities.

The more challenging question is what happens after the scan? Frontier AI models are about to hand every security team a far larger set of findings than they have ever had to work through. The real test for chief information security officers (CISOs) and IT security leaders is not how fast they can patch. It is whether they can keep the balance between patching quickly and patching correctly, at a scale no human review process was built for.

This shift creates both new challenges and new opportunities. The same AI capabilities accelerating cyberattackers are also enabling defenders to identify exposures earlier, automate remediation, and build more resilient security programs.

Not every vulnerability will be patched in time, so the controls that limit what an attacker can do by defense in depth matter more than ever. CISOs can dramatically improve the security posture of their Microsoft infrastructure by deploying Microsoft Baseline Security Mode (BSM), building on how Microsoft protects Microsoft.

How do frontier AI models change vulnerability management for CISOs?

Microsoft uses frontier AI models to find vulnerabilities in its code base and mitigate them at controlled pace. Potential vulnerabilities are reviewed on validity, severity, and potential impact. As we have explained in previous blogs, many steps in our vulnerability handling and disclosure processes now are AI-powered, allowing them to scale.

Most vulnerabilities in cloud software are being mitigated by Microsoft without customer intervention. For on-premises Microsoft software, customers should continue to expect a substantial increase in the number of vulnerabilities released on Patch Tuesdays relative to historic volumes before the advent of frontier AI models earlier this year—September 2026 saw a record number, close to 1,000.¹

Due to the non-deterministic nature of AI models, different runs by the same model or with different models may yield different results. With better models becoming available over time, AI-powered vulnerability scanning of our existing code base and new code should become part of our standard security assurance.

We use a ‘harness’ layer around AI models in vulnerability scanning for better results. This layer controls how models access code, validate outputs, and integrate findings into triage and remediation workflows. Microsoft has expanded the use of harnesses in scanning its code bases across all the engineering groups. One of these harnesses, codename MDASH, has now also been made available to customers. In addition to AI-powered vulnerability scanning, our internal Red Teaming engagements now leverage AI, enhancing the team’s capacity to find weaknesses in controls and boosting the efficiency and speed of the operations.

What can CISOs do to mitigate the risk of vulnerabilities to their organization?

How can CISOs prepare for what’s coming, as new AI models in the hands of threat actors increase the risk and pace of cyberattacks using unknown or unpatched vulnerabilities?

1. CISOs should increase the resources allocated to Microsoft on-premises software patching, prioritization, and timing as they should continue to expect a high volume of patches on future Patch Tuesdays, at least over the coming period.
2. CISOs should rethink patch timing for their most critical systems. Traditionally, critical components such as domain controllers and edge devices have been patched when downtime is least disruptive, such as weekends or holidays. As AI shortens the time between patch release and exploitation, that trade-off may need to change. Consider deploying fixes to these systems within 24 hours rather than waiting for the next maintenance window.
3. CISOs should use harnesses to scan and remediate vulnerabilities in the code base of their organizations. They should do so without delay, rather than wait for access to frontier AI models. They should allocate appropriate resources for tokens and human resources to triage and remediate bugs.
4. CISOs should focus, more than ever, on defense-in-depth and monitoring of the state of health of their critical controls. As governments and regulators around the world raise expectations for cyber resilience through legislative frameworks, these authorities are compelling organizations to strengthen their security posture, operational resilience and regulatory readiness. Extensive guidance can be found in a new Microsoft Security Exposure Management page with capabilities customers can use to act.

Microsoft helps by tackling open-source vulnerabilities with industry peers

Many organizations use open-source software in their infrastructure or their products. As governments and regulators increase expectations around software security, the ability to find and fix open-source vulnerabilities is becoming more important, but many open-source maintainers don’t have the tools or resources to respond quickly, which increases supply chain risk in the face of AI-supported vulnerability discovery.

Microsoft has teamed up with industry peers to coordinate the scanning and patching of vulnerabilities in critical open-source components before cyberattackers find them. Participating organizations are pooling resources to prioritize and scan open-source packages and remediate in cooperation with the maintainers.

Microsoft helps by offering Secure by Design and Secure by Default to customers

The implementation of defense in depth by organizations worldwide can be vastly improved by implementing baseline security controls by default across their estate. The implementation of Secure by Default requires to carefully balance useful features and fast innovation with security controls and guardrails. Microsoft is committed to implementing Secure by Design and Secure by Default across its products, thereby removing part of the burden of implementing critical controls from customers. See our latest Secure Future Initiative (SFI) report for more details.

With Secure by Design, security comes first when designing a product or service. Customers do not need to opt in and cannot opt out. Recent examples where we have implemented Secure by Design changes in Microsoft products are:

• Mandatory multifactor authentication for Microsoft Azure administrators.²
• Enforcing Conditional Access policies to Windows Hello for Business and macOS Platform single sign-on (SSO) registration.³
• Secure-by-design under the hood in Azure.⁴

Secure by Default means that security protections and secure configurations are enabled by default, requiring no extra effort. Recent examples where we have implemented Secure by Default changes in Microsoft products are:

• Azure Backup soft delete enabled by default.⁵
• Azure VNet default outbound access disabled.⁶

Customers can opt out of certain controls if they accept a higher risk posture. Customers may also create a better risk posture by opting in to more demanding or more sophisticated controls.

CISOs can dramatically improve the default security posture of their Microsoft infrastructure by deploying Microsoft Baseline Security Mode (BSM). BSM helps organizations to implement and monitor secure configurations at scale, based on “How Microsoft protects Microsoft.” Scenario analyses are available to assess the impact of additional controls in existing tenants, allowing a staged and controlled rollout. Controls can be turned on and off, and exceptions can be made and removed.

Microsoft Baseline Security Mode is available to existing customers within their current license agreement. We highly recommend CISOs to use BSM to increase the defense-in-depth of their existing infrastructure. New Microsoft customer tenants will come with BSM controls gradually implemented by default.

Final notes

The advent of frontier AI models to discover and exploit vulnerabilities creates both risks and opportunities. Responsible use of these tools allows us to secure the software that we built in the past and mitigate exposure in the future in a way not possible until now.

Microsoft will continue to use the latest technologies to find and correct vulnerabilities in existing code and to avoid vulnerabilities in future code, to decrease the likelihood of exploitation. It will also extensively use AI to protect its own infrastructure and the infrastructure of its clients.

For CISOs, the bottom line remains that risk-based management of cybersecurity is the cornerstone of our defense. AI and the principles of Secure by Default have become an integral part of this approach.

To learn more about Microsoft Security solutions, visit our website. Bookmark the Security blog to keep up with our expert coverage on security matters. Also, follow us on LinkedIn (Microsoft Security) and X (@MSFTSecurity) for the latest news and updates on cybersecurity.

¹September 2026 Security Updates, Microsoft.

²Azure mandatory multifactor authentication: Phase 2 starting in October 2025, Microsoft Azure blog. September 5, 2025.

³Protect security info registration with Conditional Access policy, Microsoft Learn.

⁴Azure IaaS: Defense in depth built on secure-by-design principles, Microsoft Azure blog. May 4, 2026.

⁵Secure by Default with Soft Delete for Azure Backup, Microsoft Learn.

⁶Default Outbound Access in Azure – Azure Virtual Network, Microsoft Learn.

The post CISO perspectives on managing vulnerability risks in the age of AI appeared first on Microsoft Security Blog.