In an industrial network, malicious activity can be hard to tell apart from normal data exchange between devices. Malicious actors routinely rely on legitimate tools and the Living-off-the-Land technique, so a single network connection may show no known signs of an attack: the device uses a familiar protocol and transmits ordinary data. An anomaly can hide in subtle details: the data source, the time of access, or the volume of transferred data may be atypical for a particular device. However, to detect this kind of anomaly, you need to understand the roles of the devices at both ends of the connection and know which network activity patterns are typical for them.

But an organization’s asset inventory data is often incomplete, which makes it harder to catch unauthorized connections in time. Suppose monitoring detects a host that regularly exchanges data with programmable logic controllers (PLCs). Its IP address is known, but the inventory does not specify what role it plays: is it an engineering workstation, a process data server, or a piece of contractor’s equipment? The host may also have been on the industrial network for a long time, with its record going out of date after a network upgrade.

In the joint OT asset inventory guidance released by CISA and its partners in 2025, inventorying devices and classifying them by role are treated as part of OT asset management. For organizations, this is one of the most important areas: according to the SANS 2025 ICS security survey, 50% of respondents had deployed or significantly expanded ICS inventory and visibility tools over the previous 12 months, and 54% named them as investment priorities for the next 12–24 months.

Device inventory in Kaspersky Industrial CyberSecurity for Networks

In Kaspersky Industrial CyberSecurity for Networks, the Asset Discovery technology is the foundation of asset visibility. Passive network traffic analysis detects devices and their interactions, while recognition rules help determine the device category and extract available information about it, such as the vendor, model, or software version. Exactly which attributes the system can obtain depends on the protocol and the traffic it observes. We covered this in more detail at the Kaspersky Industrial Cybersecurity Conference.

The product also supports active polling modules (APMs): where polling is allowed, they can request additional configuration and firmware version details from devices. This kind of inventory is far more detailed than a list of IP addresses.

However, even accurate hardware information doesn’t always explain what function a host performs in a specific industrial facility. Gaps appear for various reasons. Some devices send identifying information only at startup or during configuration, while others rarely communicate. Proprietary protocols can’t always be parsed, and encryption may leave only connection metadata for analysis. If the monitoring system doesn’t receive traffic from every segment of the industrial network, some network activity goes unobserved. Active polling doesn’t suit every device or every operating condition.

When there isn’t enough direct information about a device, its category can be refined by looking at the patterns of its recurring network interactions. Analyzing these interactions is the job of Machine Learning Event Processor (MLEP), a service included in Kaspersky Industrial CyberSecurity for Networks 5.0. This machine-learning-based event processor can both refine device categories and detect anomalies in their network behavior.

What data does MLEP analyze?

To compare device behavior, you don’t need to store every packet as a separate observation (a data object) in the ML model. MLEP represents network sessions as a sequence of short 15-second fragments. For each fragment, MLEP stores information about the endpoints (IP, MAC, and port), the direction of data exchange, the protocol, and the number and size of the packets transmitted. Together these make up an event, which then gets a timestamp. Traffic thus becomes a sequence of events showing which hosts a device communicated with, when, and how communication intensity changed.

Aggregating data into events also sets the method’s limits. An aggregated fragment can’t be used to reconstruct the exact order of individual packets or the content of commands. MLEP can draw conclusions about a new protocol only if another KICS technology has managed to recognize it. That said, comparing stable network connections and changes in exchange volume doesn’t require a detailed record of every packet. These observations form the basis of the device’s behavioral history.

From observations to MLEP results

Each object has its own behavioral history, and one object’s history can’t be applied to another. But the principles of behavior analysis, along with some common rules, work across different objects.

Let’s look at how this works in MLEP, which is built on a spiking neural semantic network (NSN).

Event identification

Neurosemantic networks structure information as a multilayer ordered graph, where the data in each consecutive layer builds on the data in the previous one. The nodes of this graph — that is, the units of information in each layer — act as neurons.The NSN creates dedicated channels to handle different event parameters (port, IP, and so on). Unique parameter values for each channel are stored as neurons in layer 0. The event itself, which consists of a set of parameter values, is stored as a neuron in layer 1 of the network. If an event occurs for the first time, neurogenesis creates a new neuron with input connections to the corresponding parameter neurons. This lets the network learn and operate on the event stream at the same time. If the event has occurred before, the network activates the same neuron that was created when the event first appeared. This way, the network identifies each event precisely. Each neuron that corresponds to an event also collects basic statistics: when it was last activated and how many times it has been activated.
 

Event history

In NSN-based event processing, layer 2 and above store sequences of events over time: recurring patterns and the full history. However, categorization and statistical anomaly detection don’t require storing the full packet sequence or a detailed timeline of every network event. As we noted earlier, events are themselves aggregates, which means the order of network packets has already been averaged out. So MLEP keeps history only temporarily, within each episode (30,000 aggregated events by default), which saves computing resources.

Statistics processing in MLEP

MLEP processes statistics mainly through specialized listener neurons on layer −1 (minus one) of the network. Attention heads and monitors are built on these neurons.

When an event arrives, the corresponding neuron is activated, and the network instantly knows whether that neuron is new or was created earlier. Each attention head subscribes to the events or event features it needs and takes in only the activations of the matching neurons. If an event occurs for the first time, all heads are notified and then decide whether to subscribe (grow a connection to the relevant neuron) and receive information about activations of such neurons from then on.

It is called a spiking network because activation information is a spike, not a number. This spike tells the head that something it’s interested in has happened.

An attention head is configured according to scenarios — the situations that MLEP should send alerts about. The core of a head’s configuration is a query, which filters incoming information and serves as the subscription criterion. Unlike regular database queries, this query runs on the stream, so it can react to an event instantly.

Attention heads work very much like the semantics of ordinary human language. A query has a subject of attention (one or more grammatical subjects) and conditions (the other parts of the sentence). In the infrastructure of an industrial enterprise, the subjects are devices on the network, and the conditions are various aspects of network communication, such as ports, protocol, and average packet size. If a head has many subjects (for example, if it’s interested in every device on an industrial network), it fans out: it creates a child head for each individual value of the subject of attention, that is, each device. The child head then watches a specific device and reports to the parent head.

Attention heads don’t accumulate or process statistics. They only filter event streams. When an attention head receives a relevant event, it activates and sends a spike to the monitor subscribed to that head.

Monitors, in turn, have stacks for accumulating and processing events. They can also have their own queries and post-filter information from the head. For example, they can select only new events from those filtered by the head.

There are several types of monitors:

• Stat monitors track changes in the average count of specific events (for example, device activity) over time intervals typical for those events, or in the average values of numeric event parameters (for example, changes in device traffic volume).
• Monitors work from the event alone: they fire when an event has occurred a certain number of times within a given interval, or has occurred for the first time.
• Similar-behavior monitors work with heads that can generalize events, that is, extract just a few parameters whose combination is shared (generalized) across many events.

MLEP uses these monitor types to implement scenarios for categorizing devices and detecting their anomalous behavior:

1. For categorization, MLEP uses attention heads and monitors that identify common behavior across different assets and group devices accordingly. MLEP then takes information from Asset Discovery about the known categories of some assets and extends those categories to other assets in the same group. Because a device can show different types of activity, it may land in several groups. MLEP then makes a weighted assessment across all relevant groups.
   

   Groups of devices with similar network behavior

   At the final stage, the system verifies that assigning the category to the asset is permissible. This is again done by special rule monitors loaded into the NSN at system startup. The rules rely on asset metadata available from Asset Discovery and APMs.

2. MLEP takes a similar approach to anomaly detection: an attention head and a monitor define an anomaly scenario, often loosely. When a monitor activates on such a scenario, MLEP generates an alert about unusual behavior.

The current version of MLEP ships with preconfigured heads and monitors.

The functional dynamic diagram below shows how events from devices on an industrial network are processed.

MLEP in practice: how behavior helps complete the asset inventory

At a facility with several similar production lines, device information may be more complete on some lines than on others. On one line, devices are identified as PLCs, while on another, after an upgrade, some hosts show up only as addresses: the available traffic didn’t provide enough information to determine their category. Still, these devices communicate with the same types of HMIs and process data collection servers. Their communication directions, frequency, and volumes may be similar.

That’s where MLEP comes in. It takes the categories of already known assets and matches recurring features of their network behavior against those of uncategorized devices. When the observed interactions turn out to be similar, MLEP can add a category for that host to the inventory.

Once the category is refined, the industrial network map provides more context about the equipment on the network. When analyzing network activity, the analyst immediately sees the device’s suggested category and can skip looking up each IP address in the design diagrams. This approach is especially useful where equipment changes faster than the documentation is updated.

The suggested category is also checked against existing device information and defined constraints. This helps filter out some contradictory options, but prediction quality still depends on how much of the traffic is visible, especially if the device communicates only rarely.

Anomalous activity and streaming statistics

Investigations of real-world attacks show that attackers actively use both standard tools and permitted protocols. You don’t always need to know an attacker’s specific technique in advance to spot potentially malicious activity. Signature-based detection helps find known signs of an attack, but it doesn’t assess how typical an observed interaction is for a particular device. MLEP compares current activity with the device’s usual network profile: which hosts it communicates with, which protocols it uses, when and how much data it transmits, its average packet size, and so on. The changes it detects help guide the direction of the investigation. Correlating these changes with other events and the operational context helps determine whether malicious actions could be behind them. When investigating an incident that has already been detected, anomalies help reconstruct the sequence of actions.

The first version of MLEP comes with six preconfigured monitors, one for each type of deviation from the device’s usual network profile:

Anomaly	How it helps with event analysis
New external host	Flags a new external connection that wasn’t previously seen in the device’s observed behavior, even if the IP address isn’t listed among known IoCs. Helps check whether this connection appeared at the same time as other suspicious events.
New combination of destination address and port	Shows that the device has begun using a service it hasn’t used before. Once an incident has been detected, this helps trace new connections after the more critical events have been analyzed.
Unusually large volume of data sent	Highlights data transfers that are atypical for the device and helps correlate their timing and direction with other events.
Atypical activity time	Shows that the device was active at a time that’s unusual for it. In an already detected incident, this helps reconstruct the timeline of actions around the main alert.
New protocol	Helps spot a new method of communication, even when the device is communicating with familiar hosts.
Activity with signs of scanning	Shows that the device has started contacting an unusually large number of neighboring hosts. This signal helps reveal possible reconnaissance ahead of attempts to access equipment.

During an investigation, anomalies help reconstruct the sequence of changes: when a specific host established new connections, switched protocols, increased its data transfer volume, or shifted its operating hours. Matching events by device and time shows an analyst how activity unfolded around an incident that has already been detected.

This is especially useful when some of the attacker’s actions fell outside the monitored scope or didn’t trigger other KICS technologies. Correlating these changes with the device’s role, protocol analysis data, and events from other technologies gives the analyst a more complete timeline. Behavioral anomalies show exactly where to look for the missing links and which network interactions need attention.

In the event list, anomalies detected by MLEP are marked with an AI label. Each event shows its name, source, destination, timestamp, and the ID of the triggered scenario.

Conclusion

In an industrial network, whether behavior is anomalous depends on the host where it occurs and that host’s place in the industrial process. The repetitive nature of many communications makes a change in normal data exchange easier to spot, while knowing a device’s function and criticality helps gauge the possible consequences. Taken together, these observations make network traffic more than a list of connections: it helps assess new events in the context of a specific facility, even without an attack scenario known in advance. Machine learning technologies are well suited to this kind of comprehensive analysis, as they can process and organize huge streams of heterogeneous data. Starting with version 5.0, Kaspersky Industrial CyberSecurity for Networks handles this function with MLEP, which works effectively alongside the product’s existing functionality.

Introduction

As part of proactive threat monitoring, Kaspersky’s Cyber Threat Intelligence team identified a malicious campaign involving C2 agents controlled by artificial intelligence. Confirmed victims of the campaign included organizations in the education sector in several countries across Africa and Latin America. CyberStrikeAI played a key role in the post-exploitation phase. This is an open-source framework in which an LLM orchestrator on the command-and-control server plans actions, selects tools, and generates commands, while an implant executes them on compromised hosts. Using CyberStrikeAI, the attackers progressed from reconnaissance to full domain compromise.

According to our data, this is one of the first public descriptions of a compromise involving an LLM in the wild. The campaign we discovered clearly illustrates how the role of AI in attackers’ arsenals is evolving: until recently, language models were used primarily during preparatory stages – for writing phishing emails, debugging code, or searching for vulnerabilities. Now, however, they are being embedded into the very core of the attack, automating actions at every stage: payload execution, persistence on the host, discovery, lateral movement, and, finally, data collection and exfiltration. Post-exploitation – the stage that traditionally required manual work by an experienced operator – plays a particularly significant role here.

A key feature of the scheme using the CyberStrikeAI framework is that the implant ceases to be a simple tool for executing the operator’s commands. Decisions are distributed across two levels: the operator defines the task in general terms, while the model determines how to execute it – specifically, which commands to send to the host and in what sequence. Thus, the strategic planning remains with the human, while the choice of technical steps is delegated to the model, and the implant executes the model’s decisions.

To understand how this scheme works in practice, let’s examine the chain of actions taken by the attackers, as reconstructed by our analysts during the incident investigation.

Timeline of the attack

Let’s examine the timeline of the attack using an example from one incident within the detected campaign.

The first phase follows a standard approach: a human operator gains access to the victim’s infrastructure via a compromised SQL Server (version 16.0.1000), then collects basic information about the host, deploys tools, and conducts network reconnaissance. All of this activity takes nearly a day, followed by a two-week pause.

The second phase is fundamentally different. The first step involves deploying the CyberStrikeAI C2 agent, which receives commands from a server-based LLM orchestrator. During the second phase, the agent performs the following:

• establishing persistence and conducting local reconnaissance within 2 minutes;
• conducting domain reconnaissance over the next 5 hours;
• after a brief pause of several days, moving on to the final stage – lateral movement via SMB and WMI, mass credential dumping on hosts, and database export – with the entire operation taking no more than one hour.

We’ll discuss these stages in more detail in the next section.

Technical details

CyberStrikeAI overview

CyberStrikeAI is an open-source, modular platform written in Go for automating penetration testing. Its key feature is combining LLM-based attack orchestration and a built-in C2 server with agents (beacons) within a single server process.

The entire “intelligent” layer is isolated on the control server. C2 functions (calling reconnaissance and exploitation utilities, managing webshell connections, assigning tasks to agents, and reading results) are available to the model as MCP tools. The implant is a classic C2 agent with no AI functionality – no API keys, no calls to AI service APIs, and no autonomous logic. It periodically polls the listener server via HTTP/S, TCP, or WebSocket, executes shell commands, and interacts with the file system. To the victim, this activity looks like a typical remote-control session, where a server-side LLM orchestrator queues tasks instead of a live operator. The LLM influences how decisions are made, but not how the corresponding commands are delivered to the host.

Although CyberStrikeAI is publicly available, specific instances of the framework can be linked to specific attackers or campaigns. Detection and attribution are possible based on the following artifacts:

• Control panel: Upon launch, a self-signed TLS certificate is generated, with the framework’s name appearing in the issuer and subject fields. The certificate is created in memory at each startup, so the not_before field effectively records the time the process was launched. Accordingly, as with Cobalt Strike, the chronology of restarts can be reconstructed from the scan history.
  

  CyberStrikeAI control panel’s self-signed TLS certificate

• Implant: The build metadata (go version -m) contains a path to the main package, which includes data pointing to the built-in CyberStrikeAI generator – a temporary directory with the build identifier. This path is identical for all payloads, making it a reliable signature. Additionally, the build info field contains version control system metadata, which can be used to compare samples against one another.
• Authentication: When communicating with the server, the implant sends an X-Implant-Token header with a fixed value in the HTTP request. The token is unique to each C2 server and is embedded into the malicious payload during the build phase; therefore, it serves as both a network indicator and a sign that the samples belong to the same server.

It should be noted that the implant version does not reflect the server version. The implant can operate for months while the operator updates the server-side components: listener configurations and keys are preserved during updates. In the incidents we detected, we observed the use of a single agent version throughout the entire incident. In such cases, the sample’s build date may indicate the earliest possible deployment date of the CyberStrikeAI infrastructure, but it still does not provide an estimate of the compromise duration.

Execution & Privilege Escalation

The first malicious commands on the victims’ systems in the campaign under investigation originated from the sqlservr.exe process, indicating the likely initial access vector and the context in which the attackers gained the ability to execute code. It is important to note that the commands marked as having been executed prior to agent installation were executed manually by the attacker specifically within the context of this process. At this stage, we are not yet dealing with automated implant activity, but rather with manual actions by the operator.

Before launching the agent on the machines, the attackers checked for its presence using a PowerShell script, from which we can deduce its default deployment paths: C:\Windows\Temp\, C:\Windows\, C:\ProgramData\, C:\Temp. At the same time, the attacker can disguise the agent’s name as that of a legitimate system process. In the example below, it mimics svchost.exe (svchosts.exe)

Like most C2 agents for Windows, CyberStrikeAI uses PowerShell to execute its payload. When a request is received from the operator, the commands to be executed are additionally encoded in Base64 to bypass keyword-based detection.

Decoded form of the command:

When administrative privileges were available (for example, a domain administrator account), instead of executing commands directly, the agent on remote hosts often created scheduled tasks so that the malicious payload would run with SYSTEM privileges:

An alternative method that does not use schtasks utility is to create a task via the Schedule.Service COM object:

The payload (p1.ps1) from the example above is a reverse shell that connects to the server at 185.186.245[.]144:5501:

The p1.ps1 payload and similar scripts were distributed across the network using shared network administrative folders and the WMI protocol, which we will describe in more detail in the Lateral Movement section.

Stealth

In addition to encoding commands in Base64 and obfuscating SQL/LDAP queries – which we will describe in the Collection & Exfiltration and Discovery sections, respectively – the attackers split the malicious payload into multiple files. It is noteworthy that the scripts were transmitted in this manner even before the agent was installed – that is, at a stage where human intervention is typically required – yet they clearly bear the hallmarks of being generated by an LLM.

Below is a sequential record of several script fragments written to the file d3.b64.

Subsequently, the attackers used the certutil utility to decode the resulting file and executed it with the following command:

Once assembled and decoded, the script’s contents looked like this:

The script is a tool for decrypting DPAPI secrets from the file C:\ProgramData\dsn_line.txt. This file, in turn, is populated by the following command, which searches for a DSN string containing connection settings in the SQL Server Reporting Services configuration file rsreportserver.config and copies its contents to dsn_line.txt.

It is worth noting that some of the scripts delivered and executed as described above contain comments in Chinese (##  –  researcher’s comments).

Persistence

If the CyberStrikeAI implant was successfully delivered, the first step was to establish persistence on the target system. To do this, the agent created a scheduled task named SystemUpdate, which specified the path to the implant as the action to be performed:

Analysis of the framework’s source code revealed an alternative method of persistence: by adding a value to the registry key HKCU\Software\Microsoft\Windows\CurrentVersion\Run, where the name of the created value matches the task name (SystemUpdate).

Agent persistence on Linux systems is also implemented in two ways – by adding the entry @reboot <path_to_beacon> & to the user’s crontab (in which case the process runs at system startup), or by adding the line (nohup <path_to_beacon> &>/dev/null &) to ~/.bashrc (in which case the agent starts when an interactive bash shell is opened).

It is worth noting that the choice of a specific persistence method is delegated to the LLM connected to CyberStrikeAI. However, thanks to the human-in-the-loop (HITL) mechanism, the operator retains full control over the process: commands generated by the model may require manual approval or correction before they are executed.

Discovery

Local Discovery

Before installing the agent on a compromised host, the attackers manually gathered basic system information using a fairly trivial set of commands:

The attackers also manually checked for internet access and obtained the external IP addresses of the compromised machines using the Chinese service dnslog[.]cn. The service generates a random third-level domain, after which the attackers attempt to resolve it on the compromised machine.

If the machine has internet access, the request succeeds, and the attackers obtain the external IP address of the compromised device on the service’s page.

For reconnaissance purposes, the agent generates a more complex script designed to map out a path for further privilege escalation by compromising SCCM or exploiting msiexec (the AlwaysInstallElevated key):

Additionally, to find ways to elevate privileges locally, the agent executed scripts to inventory services running under domain accounts and tasks running under the SYSTEM account:

During their activity, the attackers periodically checked for Microsoft Defender detections:

Remote Discovery

After gathering information about the compromised host, the attackers proceeded to conduct reconnaissance within the network. During the preparatory phase of the attack, they manually collected additional information about remote hosts using the net.exe utility.

During the main phase, the agent used PowerShell’s built-in functionality to detect accessible hosts without relying on additional utilities (such as nmap or fscan):

Additionally, relevant LDAP queries to Active Directory objects helped build a network map:

Furthermore, the presence of the installed agent on remote hosts was verified using WMI:

Domain Discovery

The next step was to gather information about the victims’ domains. To this end, before installing the agent, the attackers used standard OS utilities:

However, later on, using the CyberStrikeAI implant, reconnaissance was carried out via generated PowerShell scripts:

In addition to basic domain information, the agent performed LDAP queries to gather information for Kerberoasting and AS-REP Roasting attacks:

Finally, the agent was used to check the domain for the presence of LAPS and Microsoft Azure:

To detect signs of Microsoft Azure usage within the infrastructure, the attackers also searched for the .azcopy directory – created by the AzCopy (Microsoft Azure Storage Copy) utility – in users’ home directories:

Credential Access

After successfully conducting reconnaissance and mapping the network, the attackers moved on to the next stage – dumping credentials. For this purpose, the agent used a classic set of LOLBINs and well-known third-party tools.

The agent dumped registry hives containing credentials using the reg.exe utility:

The agent dumped the lsass.exe process using the MiniDump function of the comsvcs.dll library. In doing so, the target process’s PID was used instead of its name to make static detection based on the command line more difficult:

After successfully compromising privileged domain accounts (specifically, the built-in domain administrator), the attackers also carried out DCSync and Golden Ticket attacks via the agent using a renamed version of the mimikatz utility:

Lateral Movement

To spread across the network using compromised accounts, the attackers employed utilities from the Impacket toolkit, specifically WMIExec, as indicated by the utility’s characteristic command line:

Once installed, the agent prefers to use standard OS mechanisms, such as shared network folders.

Another method the agent uses to remotely execute commands is the WMI protocol. In particular, the agent used it to distribute the aforementioned p1.ps1 reverse shell script. This involved using compromised user accounts with local administrator privileges on the target hosts:

A second method of lateral movement via WMI uses the Invoke-CimMethod cmdlet:

Command & Control

To deliver the aforementioned tools (malicious scripts, mimikatz, and the agent itself) to the initial nodes in the infection chain, the attackers used standard OS utilities:

After the agent is installed, it checks the availability of the C2 server’s ports. The implementation mechanisms here are virtually unlimited and depend on the “creativity” of the respective LLM agents connected to the server:

To download additional utilities, the CyberStrikeAI implant also used built-in utilities (LOLBINs). The tools could be hosted on the attackers’ own servers or downloaded directly from GitHub:

Collection & Exfiltration

The data collection phase deserves special attention. It demonstrates yet another capability of LLM orchestration – the automatic generation of SQL and LDAP queries tailored to a specific task.

In one incident during the campaign, the attackers deployed the CyberStrikeAI implant on an educational organization’s SQL server and collected information from the database via PowerShell. Here is an example of a command executed by the agent:

Normalized view of the executed script with comments:

Before accessing the data, the agent checked the privileges of the account under which it connected to the database:

In decoded form, the check looks as follows:

On network resources, file collection was preceded by a recursive directory scan to identify files of interest.

To interact with the file system, the implant provides a built-in interface that supports direct uploading and downloading of files. Files are transferred directly via the main encrypted communication channel with the C2 server (HTTP or TCP), and the data itself is encoded in Base64 and packaged into the agent’s standard JSON responses.

Another built-in capability of the implant is taking screenshots of the infected machine. As with other actions, the implant does not initiate this on its own: the command to capture the screen is issued by the LLM orchestrator, and the agent merely executes it and returns the result via the same control channel.

On Windows, the command for taking screenshots looks like this:

The Linux version looks like this:

In the incidents we detected, data collection was the final stage of the attack: the chain ended there, and we observed no further activity.

Attribution

The key artifact for attribution in this campaign is the implant’s build metadata. The attackers built the malicious payload inside a Git clone of the project, so the Go compiler embedded version control system information into the binary files – the hash and commit time, as well as the vcs.modified=true flag (analogous to the +dirty suffix) – indicating uncommitted local changes in the working copy. The combination of these fields forms a unique fingerprint for a specific build.

Based on this fingerprint, the samples are divided into two independent groups, in each of which the Windows implant and the Linux implant have completely identical build metadata. A match indicates that, within each group, both payloads were built from the same project working copy.

However, these groups do not overlap with one another. They do not share a common C2 infrastructure, and different versions of CyberStrikeAI are deployed on the command-and-control servers. In other words, the build metadata reliably links the samples within each group but does not provide grounds for linking the two groups together.

The available data does not allow us to link the activity to any known group of attackers, nor to state with certainty that the same operator is behind both campaigns. Two factors prevent this. First, the tool’s openness: CyberStrikeAI is publicly available, so different operators could have independently deployed it and assembled malicious payloads under similar conditions. Second, command generation by the model: the “signature” characteristic of a specific operator is largely shaped not by a human but by an LLM. On the one hand, this increases the diversity of commands: the same command, when generated multiple times, can look different. On the other hand, it strips them of their individual characteristics, which reduces the value of behavioral metrics for attribution.

However, the Chinese-language comments found in the attackers’ scripts and the use of Chinese services still suggest a connection between the activity and a Chinese-speaking operator. In addition, analysts noted the timing of the control servers’ restarts: since the self-signed TLS certificate is regenerated each time the process starts, its not_before field records the restart time, and most of these restarts occurred during business hours in the UTC+8 time zone.

The origin of the tool itself also indirectly points in the same direction: CyberStrikeAI was developed within the Chinese cybersecurity community and is well known there.

In summary

CyberStrikeAI is a telling – though not unique – example of a shift occurring within the class of C2 frameworks: the LLM orchestrator is no longer merely an add-on to reconnaissance but has become a control layer over post-exploitation. If HexStrike AI and similar tools cover the “find and break” phase, and Sliver/Mythic/Havoc cover the “establish a foothold and control” phase, CyberStrikeAI completes the kill chain: a single server-side process written in Go executes utilities via MCP, handles webshell connections, and assigns tasks to the implant, while the model analyzes the results and decides what to do next.

The key takeaway is that the introduction of AI into the attack chain does not yet present fundamentally new challenges for detection. All the “intelligent” work remains on the attacker’s side, while the compromised host is subjected to familiar actions: from reconnaissance via Kerberoasting and AS-REP Roasting to LSASS dumps, DCSync, and the Golden Ticket attacks. This is a typical, well-documented set of domain compromise techniques. The novelty of the attack lies not in the methods themselves, but in the speed and ease with which they were executed: AI orchestration has lowered the skill requirements for the operator, but has not taken the attack beyond what defenders already know and are capable of detecting. The timing is also telling: individual key phases of the activity (such as establishing a foothold, lateral movement, and obtaining credentials) took only a few hours, and the entire attack – from agent deployment to the compromise of domain administrators – was completed within 3–4 days.

However, the CyberStrikeAI architecture does not limit users to the scenario described. The same mechanisms allow for the on-demand assembly of more complex chains – including DPAPI secret extraction and attacks on SCCM and Azure IMDS – all of which are within reach of the same group of actors, with a relatively low skill barrier. As the models evolve, the gap between the complexity of orchestration and the level of techniques used will narrow, and the operator’s role will gradually shift from setting tasks to monitoring the model’s actions.

Detection by Kaspersky solutions

It may seem that attacks using AI agents require the creation of fundamentally new detection rules. Indeed, to detect the activity of an LLM agent installed on a host within the infrastructure, it is necessary to create detection logic based on the specific characteristics of the agents. Detection rules for this activity can be found using Hunt Hub with the following filter.

However, in the case of the attack described, a different logic applies. The main difference lies in the architecture of the attack itself: the LLM agent was not located within the corporate network perimeter. It operated remotely via a C2 agent, orchestrating actions, and left no traces of its presence in the infrastructure (such as specific API requests to local LLMs or anomalous patterns of resource consumption by hosts).

The key difference between this activity and an attack by an experienced human operator lies largely in speed. Tactically, technically, and methodologically, AI, for the most part, uses exactly the same tools, utilities, and techniques as a human penetration tester. It does not write unique exploits from scratch, but rather combines known methods and dynamically adapts them to the situation, handling the tools’ errors on its own.

Since the actions performed by an AI agent on the network are indistinguishable from those of a human, classic, well-tuned correlation and behavioral analysis rules are sufficient to detect it. Here are a few examples of how Kaspersky Next EDR Expert successfully detects the attack described:

• To detect the execution of a malicious script via a remote scheduled task, you need to monitor the creation of the schtasks.exe process with the /s argument. EDR Expert detects this activity using the rules create_task_on_remote_host_using_standard_tool, scheduled_task_creation_as_system_via_schtasks, and scheduled_task_with_powershell_execution.
  
• During the attack, the adversary will need access to the victim’s credentials; in the described case, one way to obtain them was through a well-known technique that uses legitimate tools to dump LSASS, specifically by calling the MiniDump function from comsvcs.dll. The EDR Expert solution detects this activity using the rules dumping_memory_with_com_services_dll and suspicious_lsass_memory_access.
  
• A significant part of the attack chain is lateral movement, which helps the attacker navigate within the infrastructure and achieve their desired objectives. In the described case, the WinRM protocol is used to initiate connections directly from PowerShell. This is a standard technique that allows the attacker to execute commands on remote machines without using additional network tunnels. The EDR Expert solution detects this activity using the rules using_powershell_for_remote_execution_via_wmi and winrm_send_command.
  

Kaspersky Anti Targeted Attack (KATA) is used to detect malicious activity in network traffic. Communication between CyberStrikeAI agents and the command-and-control server occurs without accessing AI platforms such as ChatGPT or Claude. Instead, the agent’s C2 server is an infrastructure pre-configured by the attackers, just as in attacks that do not use AI agents. Analysis of decrypted network traffic using NDR-class solutions reveals characteristic communication patterns of CyberStrikeAI agent activity. When these patterns are detected, KATA generates an alert with the verdict Backdoor.CyberStrikeAI.HTTP.C&C.

To detect post-exploitation activity, you can use not only IDS technology but also NAD (Network Anomaly Detection). It analyzes all traffic for suspicious indicators that do not correspond to a node’s typical network activity.

For example, actions performed by CyberStrikeAI during the Domain Discovery phase are detected by an NDR:NAD rule named “Scanning domain records via LDAP queries.”

You can learn more about NDR:NAD technology in our article on Securelist.

You can also visualize the connections between C2 addresses and files used by attackers in the Research Graph section of the Kaspersky Threat Intelligence Portal. This includes tracking the emergence of new CyberStrikeAI implants that communicate with malicious addresses.

In addition, the Threat Lookup service available on the portal allows you to determine whether malicious files belong to CyberStrikeAI using attribution mechanisms based on the executable file’s code.

Kaspersky SIEM detects all stages of an attack by the CyberStrikeAI agent. The product repository contains over 100 rules for detecting the techniques discussed in this article; we will list the most critical ones below.

Execution of various commands from the sqlservr.exe process:

• R300_04_Windows shell started by a web application

Downloading and executing a payload via PowerShell:

• R110_03_PowerShell code downloaded and executed

Creating Windows scheduled tasks:

• R099_04_Suspicious arguments when running schtasks.exe
• R300_03_Windows shell started by the scheduler

Persisting via crontab or bashrc:

• R099_06_Suspicious cron job activity
• R405_02_UNIX shell configuration modification

Rules located at the KUMA Packages/SOC package/Configuration discovery path detect the collection of information about hosts, network settings, and domains using various reconnaissance tools. Rules that aggregate detections by reconnaissance tools:

• R224_14_Windows reconnaissance activity
• R224_09_System information discovery by running multiple utilities within a short period
• R224_04_System information discovery via standard Linux tools

Data collection for Kerberoasting and AS-REP Roasting:

• R034_01_Multiple TGS ticket requests (Kerberoasting)
• R452_18_AS-REP Roasting

Dumping registry hives containing credentials using the reg.exe utility:

• R230_02_Copying data containing credentials from the Windows registry

Dumping the lsass.exe process using the MiniDump function of the comsvcs.dll library:

• R233_04_Process memory dump via comsvcs.dll
• R262_Suspicious access to the LSASS process

DCSync and Golden Ticket attacks using the renamed mimikatz utility:

• R073_DC Sync attack
• R450_41_Golden Ticket Attack
• R100_03_Traces of Mimikatz use via cmd

Lateral movement using WMIExec:

• R460_08_Lateral movement using WMIexec Impacket
• R460_09_Lateral movement using WMIExec and Impacket

For the rules to work, you must configure the SIEM system to receive the relevant events:

• Process launches (Windows security log – 4688, auditd – execve) and PowerShell commands (4104)
• User logins to the system (Windows security log – 4624, 4768, 4769)
• Various object manipulations (Windows security log – 4662, 4663, 4656; auditd – openat)
• Interaction with shared resources (Windows security log – 5145)

Indicators of compromise

File-based indicators:

CyberStrikeAI Windows Beacon
D726A8B4886C636F76CA6D30EF6A6FF4	c:\windows\temp\svchosts.exe
2770562d104391b92b42671c270079bb	c:\windows\temp\svcmon32.exe c:\programdata\svcmon32.exe
CyberStrikeAI Linux Beacon
09e1da2afa1216ecc55648ec8e122d2a
805B9E4AE74C903787416800999B8418
Tools & Scripts
29EFD64DD3C7FE1E2B022B7AD73A1BA5	C:\Windows\Temp\wdiag2.exe (mimikatz)
74F57B3E897A4960DCDD5B7E8A54F9DF	c:\rwx\RogueWinRM.exe
0BBD08AF95F877F1D568B03DE2EB76B4	c:\ProgramData\p1.ps1 (reverse shell PS script)

Network indicators:

185.186.245[.]144	C2 for Windows Beacon
fastcdn[.]bar	C2 for Windows Beacon
media.fastcdn[.]bar	C2 for Linux Beacon
104.234.26[.]163	C2 for Payload Delivery + Linux Beacon

A modern organization protects its information systems with a whole suite of technologies that generate thousands of security alerts every day. Looking at these alerts in bulk, you’ll see that a significant share are false positives. Ideally, these alerts would be filtered out automatically so that analysts could focus on the ones that may point to a real incident. However, writing filters is a time-consuming task in itself. To create a correct filter, an analyst needs to:

• Review the logs
• Identify the source of the activity
• Define filtering boundaries
• Choose filtering criteria
• Write a rule
• Check that the filter is robust

This post looks at how filter creation itself can be automated to reduce analyst workload.

Filter validity criteria

In the context of security alerts, a filter is a combination of event field values that marks a detection as a false positive and not worth the analyst’s attention. Filters can be generated for any combination of event fields, but not every combination yields a valid filter. That’s why we need to define validity criteria: conditions showing that a specific filter actually removes noise and doesn’t suppress useful alerts.

Detection rules in security systems typically contain a set of conditions, and each of them can generate an alert. A specific event triggers a given fragment of a detection rule. The set of events that a given fragment responds to is what we’ll call its context. A good filter should cover the entire context where possible, without going beyond it. Filters that are too broad or too narrow are useless. For example, a filter based solely on PowerShell would exclude too many events, including some that deserve the analyst’s attention. Most of the filtered events would also be unrelated to the rule in question. A filter on a specific ID, on the other hand, would cover only one event, which may never recur.

Problems in filter generation

We can run into a number of problems when creating filters. First, events often contain an excessive number of fields (around 400 on average), and most of them aren’t needed for filtering. At the same time, generating and ranking combinations of these fields is also fairly resource-intensive. So, the first step is to discard the irrelevant fields and keep only those worth using in the filter. For example, out of 400 fields in an event, four may be enough for filtering.

Another problem is string normalization. Field values (for example, the path to the same application on different systems) can vary depending on system settings, user folder names, and so on. For a filter to account for all these variations, the strings must be normalized. For file paths, in particular, you can replace the drive letter and user folders with standard environment variables and convert all paths to a single case.

Finally, if a filter completely mirrors the query of a detection rule, it effectively disables that rule. To fix this, remove the event fields used in the detection rule from the filter. For example, if a rule triggers on a file path, there are two ways to filter out unwanted alerts:

• In the filter, use fields other than the file path.
• Use a part of the file path that the rule doesn’t use. For example, for the rule file_path: ".../something.exe", you can set up a filter that excludes events with the user/unique_files/ directory in the file path.

Filtering security alerts with an algorithm

Given the problems described above, automatically generating filters for security alerts should consist of the following steps:

1. An analyst receives a security alert that they assess as a false positive and uses the filter generator.
2. The generator takes the 400-field event from the alert, along with the rule that triggered on it, and determines the context.
3. The generator selects the fields that are relevant to the filter.
4. The generator normalizes the strings.
5. The generator builds different combinations of the selected fields. To do this, we applied the FP-Growth (frequent pattern growth) algorithm to the context determined earlier. FP Growth builds a tree of frequently occurring items (fields) from a dataset (in our case, the context) and discards items that aren’t frequent enough. This reduces the number of combinations to check, which lowers the cost of finding the most relevant ones.
6. The generator ranks the resulting combinations and offers the analyst the top three to choose from.

For example, consider an event described by the following JSON:

After generating and ranking the combinations, we got three filters based on file paths:

• processcmdline:("nodejs\\node.exe" AND "*****\\appdata\\roaming\\npm\/node_modules\/pi-coding-agent\/dist\/cli.js")

• file_path:"*****/.pi\\agent\\bin\\fd.exe"

• processcmdline: "programfiles\\nodejs\\node.exe"

The last filter turned out to be too broad, but the analyst marked the other two as valid.

Pros and cons of the algorithmic approach

The algorithmic approach to filter generation has several advantages. First, it’s simple and deterministic: the same alert always produces the same result. Second, it’s interpretable, meaning we can see why a given combination ranked near the top of the list and why a particular field didn’t make it into the list of candidate fields at all. Third, the stages of the process are independent of each other, so individual steps can be improved and tuned without the risk of breaking the algorithm elsewhere. Finally, it has no external dependencies or models, which makes it easy to deploy in any customer environment.

However, it also has drawbacks. The first is its dependence on context. If there’s no context for some reason — for example, the rule’s first-ever trigger — a filter can’t be created. The second problem is string normalization. There’s a high chance of running into atypical cases that the normalization rules don’t cover. Often such cases can’t be described correctly at all, because they contradict each other.

These problems can be solved with large language models (LLMs). When an LLM generates the filters, no context is needed, because the LLM doesn’t iterate over or rank combinations. The LLM also takes a more flexible approach to string normalization, including non-standard cases. That said, an LLM doesn’t have to be involved at every stage of the generation process. We decided to keep the algorithmic solution where it works well. For example, a simple algorithm selects suitable fields without errors, so we didn’t move this stage to the model.

Filter generation with an LLM

The LLM-based filter generation looks like this:

The input is still an alert containing an event with 400 fields, plus the rule that triggered on that event. Then an algorithm selects the most suitable fields for the filter. Next, the LLM processes the input — the event, the selected fields, and the detection rule — and generates three filter options along with an explanation for them. The explanation helps the analyst understand why these particular filters were suggested and make an informed choice.

Consider the following event in JSON format:

The LLM suggested three filter options that all contain the same command line fragment:

• file_path: "\\WindowsPowerShell\\powershell.exe" AND filecmdline: "Get-VM -Name kafka\*"

• filecmdline: "Get-VM -Name kafka\*"

• filecmdline: "kafka*"

As in the previous example, two of the three output filters were valid. The last one was too broad to be usable. The analyst chose the first filter with minor edits:

file_path: "powershell.exe" AND filecmdline: "Get-VM -Name kafka\*"

Pros and cons of generating filters with an LLM

Using an LLM solved the problems of purely algorithmic filter generation: dependence on context and errors in string normalization. However, the single-request approach had its own drawbacks. First, it had no output validation. The LLM returned filters as-is, with no syntax check and no way to correct errors. Second, unlike in the algorithmic version, all filter generation tasks were handled by a single system prompt. As a result, a minor fix in one place could break the logic in another.

Using multiple AI agents

To address the problems of generating filters with a single model and a single system prompt, we tried distributing the tasks across different LLM agents. In this version, we brought in agents as early as the stage of selecting relevant fields. The chain includes four agents in total:

• Selector: selects the relevant fields.
• Generator: generates filters.
• Validator: evaluates the generated filters. If the filters are found to be invalid, it returns the task to the generator. If the selected fields are found to be irrelevant, it starts a new field selection cycle. If the filters pass validation, or the limit on regeneration attempts is reached, the valid filters (or the best ones available) are passed to the analyst.
• Commentator: analyzes the output filters and explains which field values were used and why.

For the following JSON event, the agents produce three valid filters based on the file path.

The filters differ in the additional parameters that support the detection:

• file_path:"sysvol\\domain\\scripts\\logon.bat"

• file_path:"sysvol\\domain\\scripts\\logon.bat" AND file_md5:"0x7..6E"

• autorun_entry: "Group Policy\\State" AND file_path: "SYSVOL\\domain\\scripts\\logon.bat"

Pros and cons of the agent-based approach

LLM agents give us a system that can validate and correct filters. Its modular architecture also lets us tune and improve different stages of generation independently. However, this approach makes more LLM calls than the previous ones, so the system is slower. It also depends on access to a model, which may not be available in the organization’s isolated environment.

Evaluating the filter generation system

To evaluate how well the filter generation system works, we use two main metrics:

• Approval rate. For each security alert, the system suggests three filters. If the analyst selects at least one of them, we count the generation as successful. For example, in the case below, the approval rate is 2/3.
  
• False positive reduction rate (FP drop). This metric shows how much a filter helps reduce the number of false positives. It can only be measured over time, because we need to compare the number of false positive alerts received during a set period before and after the filter is deployed.

Below is a comparison of the three filter generation approaches by these metrics:

As the results show, LLMs significantly improve the efficiency of filter generation. At the same time, the LLM agentless and agent-based approaches perform almost identically. So we chose the simpler and faster option: an LLM with a single system prompt. As a result, we cut false positives by 99%, increased the number of filters generated per day by 28%, and made analysts’ alert handling 17% more efficient.

Takeaways

Practice has shown that filter generation can be automated effectively. We’ve looked at several options for automatically generating filters for false-positive security alerts: a purely algorithmic approach, an LLM-based approach, and an approach built on a group of LLM agents. All of them can generate valid filters and make the analyst’s work much easier, but each has its own pros and cons.

The algorithmic approach scored lower on the evaluation metrics, but in two-thirds of cases, it still produces filters that satisfy the analyst. It works especially well when there’s already a set of similar events to filter out, meaning enough context has accumulated. The algorithmic generator can also be deployed in any environment. Both standalone LLM and multi-agent LLM approaches show high accuracy, and their results differ only slightly, so the slower agent-based approach looks excessive at this stage. If an LLM is available, the best option appears to be a single model that handles several stages of the generation cycle.

AI systems are much harder to protect than traditional IT — whether from operator mistakes, malicious interference, or just plain unstable behavior. There are two main reasons for this. First, AI outputs are probabilistic, so the same prompt can give different answers each time you run it. Second, today’s LLMs can’t reliably tell commands apart from data. Despite the demonstrated security issues of AI systems that we’ve covered many times, it’s impossible to ignore the potential economic payoff of adopting them. That’s why many companies are moving beyond simple AI assistants to AI agents. These agents carry out tasks on their own and have direct access to the same tools as humans — from reading databases to changing settings in production processes. Our experts, in their new study, examine how to clearly define and minimize the resulting risks to keep potential problems to a minimum. The approach borrows and adapts the principles of Cyber Immunity and Security by Design.

The basic principles of Security by Design

All the principles of this discipline are universally applicable to the design of any computer system — AI agents included. What differs is how deeply each principle needs to be adapted and reworked for AI scenarios.

1. Security is a separate and important factor right from the design stage. Security goals and assumptions are set on day one and followed throughout every stage of development. For AI agents, these assumptions should include three key points: (i) the language model is treated as untrusted, (ii) external data of any kind is considered potentially malicious, and (iii) irreversible actions can’t be performed autonomously without mandatory human approval.
2. The system is designed on the assumption that it will be attacked.
3. The trusted computing base and attack surface are kept to a minimum. Agents get only the tools they actually need, limited permissions, and the smallest amount of context possible. Security policy enforcement is handled entirely by deterministic components: the policy monitor, the agent harness, and isolation mechanisms (sandboxes).
4. The principle of least privilege applies. Each agent has its own set of credentials and access privileges (identity), limited to the bare minimum it needs to do its job. If an agent temporarily needs elevated privileges, it gets them only after external approval, and they last exactly as long as the task that requires them.
5. Components are isolated. In agent-based systems, this is one of the most complex and multi-layered architecture problems. The agent harness should run in an isolated environment (sandbox or virtual machine), and agents should be kept as isolated from each other as possible. Data processing in trusted and untrusted contexts is also kept separate. For example, one agent reads a pull request and writes a draft of the code, while another reviews that work without ever seeing the potentially malicious text of the request.
6. Default deny is the rule. Security policies follow the rule: anything not explicitly allowed is forbidden. This includes tool calls, interactions between agents, and network traffic.
7. Defense in depth is in place. Different measures should work at the same time and reinforce each other. Probabilistic defenses — like AI guardrails (a “firewall” for language models) or alignment (the model’s own sense of what’s acceptable and what isn’t) — can’t offer any formal security guarantees. Nevertheless, they make the system more secure overall and serve as a valuable source of telemetry.
8. The supply chain is protected. Some agents can install skills, libraries, and other software components on their own, so these abilities need strict rules. Either block them completely, or limit them to a closed, pre-approved list of specific versions of specific components. Systems built on cloud LLMs face an extra risk: model updates. One day, the same API request might be answered by a different language model with different security properties.

Pitfalls of AI security

The general principles of Security by Design are well known and look simple on paper. But applying them to AI systems is tricky because agents are incredibly flexible. For example, a security team might believe that an agent on a developer’s computer runs in a sandbox and can’t change anything outside it. But an attacker, or even the agent itself, can come up with dozens of ways around this seemingly solid protection: from running commands through a local MCP server that the sandbox doesn’t restrict, to reconfiguring another AI agent installed on the system. To make things even harder, the range of attack vectors is huge: from arbitrary fields in the agent’s configuration to innocent-looking text strings in a skill file.

One serious problem that designers often underestimate is user fatigue. A busy agent can throw up dozens of requests every minute, asking the user to confirm potentially dangerous actions. Within an hour, the person is clicking Approve on autopilot without reading what they’re approving. That’s why it’s important to carefully tune both how often these requests appear and how they’re worded. Nobody will read a request description that’s too long. And one that’s too short can hide the key details the user needs in order to make the right call.

Vulnerabilities in agents and related tools are a pressing problem — and it’s made worse by how quickly these categories of software evolve and receive updates. Industry-standard platforms and apps — such as LangGraph, MS Semantic Kernel, and Claude Code — have had critical vulnerabilities patched more than once. Architects should always consider swapping a large, do-it-all framework for a compact, stripped-down system that has only the features a specific agent actually needs.

On top of that, carefully implementing standard, long-established security measures in the available tools greatly reduces the damage if an agent is attacked or behaves unexpectedly. One example is a strict allowlist of domains that the web fetch tool can open when asked by the agent. But designers rarely account for the hidden problems and overhead associated with maintaining the list. As a result, it slowly goes out of date and leaves loopholes for attackers. The most striking example is the EchoLeak attack. Microsoft 365 Copilot had a properly implemented and fairly effective allowlist for accessing external sites, but attackers circumvented it through a feature in a secondary tool and the MS Teams domain.

Practical tips for architects and security teams

AI system architects and heads of engineering and security teams should build in agent security from the design stage. They should rely primarily on verifiable, deterministic mechanisms: isolating components, controlling interactions, and keeping the trusted computing base to a minimum.

Key architectural decisions include defining acceptable autonomy for the agents, limiting access to confidential data and dangerous combinations of capabilities, and using sandboxes, separate roles for different models, and access control gateways. Restrictions added as an afterthought reduce risks, but they’re no substitute for secure architecture.

At the organizational and operational levels, companies need to keep an inventory of their agents, coordinate the efforts of security and deployment teams, protect container environments, manage the supply chain (AIBOM), and integrate multi-layered defense.

Observability requires special attention. For AI systems, this should include logging the LLM’s requests, responses, and reasoning, as well as detailed telemetry from the hosts and cloud environments where the agent harness runs.

The hardest part of designing agents is balancing security, performance, and usefulness. To learn how to strike that balance with threat modeling and clearly defined security goals, read the full version of the study.

Security teams are starting to actively use AI for security work, including vulnerability triage, penetration testing, threat modeling, incident response, and code review. The promise is speed, but a security tool that moves fast and raises too many false alarms doesn’t save time. Engineers spend time on false alarms, on-call is noisier, and teams distrust findings that matter.

Today, we’re releasing Deception Benchmark, the first benchmark designed to measure that trust problem directly. It tests whether a model can distinguish real vulnerabilities from code that looks risky but is actually safe. The benchmark includes 14,822 samples across 16 languages and more than 70 Common Weakness Enumeration (CWE) categories. We evaluated 12 models from five providers and are releasing the dataset and whitepaper to the community. Existing benchmarks measure whether AI can find or exploit vulnerabilities. This is the first to measure whether it can tell real vulnerabilities from false alarms. Under standard prompting, precision at distinguishing real vulnerabilities from false alarms landed in the mid 50s; as likely to be inaccurate as accurate.

In offensive tasks, there’s often a clear result: the exploit works or it doesn’t. Defensive reviews are harder to verify than offensive tasks; a model might recognize a suspicious pattern even when a mitigation makes the issue non-exploitable. In practice, useful systems need to reason about the code, the mitigation, and sometimes the surrounding environment.

The community has made progress on security evaluations. CyberGym tests agents on more than 1,500 realistic tasks. Meta’s CyberSecEval and CyberSecEval 2 measure exploit generation. CYBENCH evaluates capture the flag (CTF) challenges. SEC-Bench and VulnBench push toward authentic security workflows.

Recent work reinforces both the progress and the gap. ExploitGym measures whether AI can escalate from a crash to a working exploit. Microsoft’s Project Perception deploys multi-agent red/blue/green teams for continuous defense. OpenAI’s GPT-Red shows that self-play red-teaming finds novel attacks that frontier models can’t defend against. Since then, OpenAI disclosed that its GPT-6 Astra model crossed the Critical cybersecurity capability threshold, and both OpenAI and Anthropic reported incidents where models gained unauthorized access to production systems during evaluations. The offensive side is moving fast. But none of this work measures the defensive precision question: when an AI system flags code as vulnerable, how often is it right?

14,822 samples, 16 languages, and more than 70 CWE categories. We call it Deception Benchmark because the safe samples are designed to deceive models. It has real vulnerability patterns, real frameworks, real idioms, with mitigations that quietly close the exploit path. The goal is to classify code as vulnerable or safe, with no hints.

Consider a Flask endpoint that accepts user input and queries a database. A model will pattern-match to SQL injection, but the query uses parameterized statements, so the exploit path is closed. A single-turn classifier flags the pattern and moves on, never checking whether the exploit can actually work. Production tools rely on multi-step loops and agentic workflows to compensate, but that scaffolding masks whether the model itself understands the code. This benchmark strips the scaffolding away and asks the model to make the call in a single pass, so what it measures is understanding, not how many tries a harness takes to get there.

We built every sample through an adversarial loop: generate, test against frontier models, harden, repeat. If a model gets it right easily, the sample doesn’t survive. The result is a benchmark calibrated to the frontier, not below it. Building it this way is expensive. Generation and hardening of the samples consumed tens of billions of tokens. We’re releasing the result so the community doesn’t have to repeat that cost.

This benchmark generates two challenge types. Code-level challenges (6,988 samples) present vulnerable and safe variants that differ by a subtle fix. Both look suspicious, only one is exploitable. Environment-gated challenges (2,707 samples) go further: same code, different deployment context. A Kubernetes Network Policy blocks the server-side request forgery (SSRF) path. An identity and access management boundary prevents privilege escalation. The pattern is visible in the source. The infrastructure makes it unexploitable. The model has to figure out which scenario applies.

All samples were purpose-built for this benchmark, grounded in real-world patterns, real frameworks, real CWEs, and real infrastructure; without IP concerns or training data contamination.

Generating reliable labels at this scale is difficult: a single pass—by people or by models—leaves errors that skew scores. So we treat labeling as a convergent audit loop rather than a one-time step. Every label is re-examined by multiple independent reviewers, blind to one another and to the original reasoning that produced the label. Disagreements escalate to direct adjudication, where the original reasoning is evaluated against the challenge. Unresolved cases go to human review. We repeat the loop until the scored set converges below a dispute threshold: under 3 percent of samples still contested by independent review, with a target of under 1 percent surviving human adjudication. One choice makes this defensible: we never relabel a disputed sample. When reviewers disagree, the sample moves to the unscored pool instead of being given a corrected label, so a bad challenge can remove a sample but can never introduce a wrong label into the scored set.

A human review of 100 randomly drawn scored samples found no label errors. We describe the full process in the whitepaper.

The benchmark is roughly balanced: half vulnerable, half safe, so a random classifier scores 50 percent. We report two error rates separately, because they fail in opposite directions. The false positive rate (FPR) is how often the model flags safe code as vulnerable. These are the false alarms that waste an engineer’s time. The false negative rate (FNR) is how often it misses a real vulnerability and calls it safe. Accuracy alone hides this: a model that labels everything vulnerable catches every bug (0 percent FNR) but flags all safe code (100 percent FPR) and still scores about 50 percent. We consider FPR below 10 percent and FNR below 10 percent the minimum bar for production use.

Model	Prompt	Accuracy	FPR	FNR
GPT-5.6 Sol	Direct	54.9%	92.5%	0.9%
GPT-5.6 Sol	PoE	58.9%	58.6%	23.1%
GPT-5.5	Direct	56.9%	87.8%	1.3%
GPT-5.5	PoE	62.9%	63.6%	12.4%
GPT-5.4	Direct	60.2%	81.0%	1.5%
GPT-5.4	PoE	77.7%	10.1%	33.6%
Llama 3.3 70B	Direct	58.8%	84.2%	1.1%
Llama 3.3 70B	PoE	72.2%	10.2%	44.2%
Claude Haiku 4.5	Direct	55.6%	92.1%	0.0%
Claude Haiku 4.5	PoE	75.6%	22.4%	26.3%
Claude Opus 4.6	Direct	55.9%	91.3%	0.1%
Claude Opus 4.6	PoE	75.8%	42.7%	7.0%
Claude Opus 4.7	Direct	58.3%	85.5%	0.9%
Claude Opus 4.7	PoE	75.9%	32.0%	16.8%
Claude Opus 4.8	Direct	53.8%	95.7%	0.2%
Claude Opus 4.8	PoE	75.8%	32.5%	16.4%
Claude Opus 5	Direct	77.3%	41.5%	5.2%
Claude Opus 5	PoE	79.3%	24.9%	16.8%
Claude Sonnet 5	Direct	62.9%	74.7%	2.2%
Claude Sonnet 5	PoE	74.7%	31.8%	19.2%
Amazon Nova 2 Lite	Direct	56.3%	89.2%	1.2%
Amazon Nova 2 Lite	PoE	70.1%	45.2%	15.5%
Mistral Large	Direct	52.2%	99.0%	0.0%
Mistral Large	PoE	65.5%	49.3%	20.6%

Among the general-purpose frontier models tested, no configuration achieves both FPR and FNR less than 10 percent on this benchmark.

Every model has the same failure mode. With direct prompting, they catch up to 95 percent of real vulnerabilities but also flag 41–99 percent of safe code. Precision runs from 52 percent to 71 percent, clustered in the mid-50s; effectively as likely to be inaccurate as accurate. The models see a vulnerability pattern and stop reasoning. Proof-of-exploit prompting cuts false positives by 17–74 points but misses 7–44 percent of real vulnerabilities. The environment-gated challenges are worse: models flag the code and ignore the Kubernetes Network Policy next to it. No tested configuration keeps both false positives and false negatives below 10 percent.

These results reflect general-purpose models in single-turn prompting. Purpose-built systems with multi-step validation and tool use are a different operating point that we didn’t measure, and if a harness can close the gap between pattern recognition and genuine understanding, this benchmark is the place to demonstrate it. Two cautions before assuming it already does. Agentic verification is proven mostly on offensive tasks, where success can be confirmed: the exploit fires or it doesn’t. Judging that code is safe has no such oracle. Extra iterations re-sample the same judgment rather than confirm a negative, and a harness still inherits the base model’s understanding. If the model can’t separate an effective mitigation from an ineffective one in a single pass, more passes won’t add the missing knowledge. That’s what this benchmark measures: the model’s intrinsic ability to understand code, tested at the single-turn baseline where no scaffolding can mask the gap.

For security teams evaluating AI tools today: ask your vendors how their system performs on tasks like this, not just whether it finds vulnerabilities, but how often it’s wrong. Pair any AI-assisted review with human verification on high-risk code paths, and use Deception Benchmark to hold your tools accountable.

We built Deception Benchmark to simplify measuring this problem in a reproducible way. The public release includes the samples and evaluation workflow. We don’t release the labels, so submissions can be scored consistently over time without turning the benchmark into a memorization exercise.

Of the 14,822 samples, 9,695 are scored; the remaining 5,127 are held out and unscored, mixed in with the rest of the benchmark. The goal is straightforward: make it more difficult to optimize the benchmark compared to improving the underlying system. We describe that design in more detail in the whitepaper.

Deception Benchmark is available on GitHub, along with the whitepaper and submission instructions for verified scoring. If you’re building security tooling, you can download the dataset, run your system against the benchmark, and submit predictions for scored evaluation.

If you have feedback about this post, submit comments in the Comments section below.

---

Cybersecurity news increasingly focuses on attackers using AI for reconnaissance, vulnerability hunting, and writing scripts or malware. Against this backdrop, a less conventional category of attacks remains largely overlooked: threat actors leveraging AI tools that are already deployed within the target infrastructure.

AI agents designed for coding, as well as command-line tools — such as Claude Code CLI, Gemini CLI, Codex CLI, Amazon Q CLI, and the like — are particularly attractive targets. These agents can read and modify files, execute shell commands, install packages, and connect to external services. Developers frequently allow them to run in autonomous mode to avoid manually confirming every action. For an attacker, such an agent becomes a ready-made multitool that has the right privileges, access to critical data, and security clearances from the get-go. That said, other types of AI agents, such as automated financial workflow assistants, are equally vulnerable to exploitation.

Why trusted agents pose a threat

An AI agent differs from standard software in its broad access privileges, the wide array of ways it can be prompted to execute actions, and the difficulty of detecting its anomalous behavior. Enterprise security controls (EPP, EDR, XDR, SIEM) scan for malicious files, known command patterns, and suspicious action sequences. Even if a specific strain of malware hasn’t been seen before, its behavior often gives it away: bulk searches for documents, spawning PowerShell processes, accessing credential stores, or exfiltrating data to an external server from an unknown application will trigger a deluge of alerts in the SOC.

It’s much harder to apply that same detection logic to a legitimate, pre-installed AI agent. By design, these tools are capable of executing an extremely broad range of operations. Running scripts, reading configuration files (including secrets), moving documents, and installing dependencies can all be a part of their normal operation — especially on a developer’s machine.

When an attacker attempts to exploit this kind of AI system, they use natural language to convey their instructions to the agent. A request to hunt for passwords, API keys, or cryptocurrency wallets can be framed in thousands of ways, expressed in a variety of languages, or phrased indirectly. Furthermore, a malicious command can be hidden in any data stream the agent reads — including text fields in documents, file names, or even error logs, as shown in one of the examples below.

Integration between agents and third-party software further expands the attack surface. Agents are frequently connected to email, instant messaging apps, task management systems, knowledge bases, and MCP servers. As a result, a malicious instruction can originate from virtually any source whose data the agent processes automatically.

Below are several examples of attacks where threat actors leveraged AI agents to achieve their goals.

The Nx attack: a coding agent hunting for passwords

The most prominent and widespread example is the compromise of Nx npm packages in August 2025, dubbed s1ngularity. Attackers exploited a vulnerable GitHub Actions workflow, stole a package-publishing token, and released trojanized versions of several @nx packages that harvested developer secrets.

The malicious post-install script checked whether Claude Code, Gemini CLI, or Amazon Q CLI were installed on the victim’s machine. If any of these were detected, the agent was issued an instruction to scan for cryptocurrency wallets, ENV files, API keys, and other sensitive data. The agents were executed with auto-approval flags enabled, such as --dangerously-skip-permissions, --yolo, and --trust-all-tools.

The search results were then exfiltrated to public GitHub repositories created by the attackers. Over several waves of the attack, thousands of secrets belonging to hundreds of organizations were exposed.

Instead of developing a complex, custom harvesting tool, the attackers repurposed the victim’s own legitimate agent. It already could navigate the file system, understand file context, and identify high-value target data. However, post-incident analysis by security experts revealed drawbacks to this approach: the AI occasionally refused to carry out unethical tasks. Furthermore, the malware authors had to adjust their code so that it “waited” for the AI agent, which could take a considerable amount of time to process the task while generating significant intermediate output.

Communication via… error logs

The AgentJacking study was not a malicious attack, but it differed from one only in intent: its authors had no interest in stealing secrets. After developing the attack technique, the researchers tested it on a global scale and received callback responses from AI agents belonging to more than a hundred real-world companies — some of them major enterprises.

The attack targeted Sentry, a widely used telemetry service. Websites and mobile applications automatically transmit error reports to Sentry so developers can monitor issues without requiring active user reporting. These reports are accepted without authentication because errors often occur for anonymous visitors that the site or application knows nothing about.

The researchers configured their own MCP server within Sentry, allowing AI agents to analyze reports automatically. Next, they submitted a fake error message containing a data block formatted to match Sentry’s MCP server data structure. Embedded inside was an indirect prompt injection — an instruction to run “additional diagnostics” via a command like: npx @controlled-validation-package --diagnose.

If error handling were delegated to an agent, it could take the instruction as legitimate and execute a command resulting in the automatic installation of a third-party package. In this experiment, the package merely contacted the researchers’ command-and-control server without exfiltrating any sensitive data. A genuine malicious package, however, could steal credentials, tamper with code, or establish persistence in the system. Identifying targets for such an attack is equally straightforward: an organization’s use of Sentry is easily detectable by analyzing website code or checking the strings in mobile application binaries.

Notably, according to the study’s authors, Sentry acknowledged the issue but stopped short of implementing comprehensive fixes, opting instead to block only the specific variant of the malicious query routed through the MCP. Screenshots included in the study reveal that a Sentry employee described the system as “technically not defensible” against the attack.

Stealing financial data via MCP

Abusing unsecured MCP servers — as seen in the Sentry case — is just one way to attack this protocol. Introduced just a year ago, and already dubbed “USB for AI agents”, the Model Context Protocol standard was adopted rapidly without robust built-in security controls. As a result, attacks targeting MCP are bound to proliferate — a trend Microsoft recently warned about.

Threats can lie hidden even within the description of an MCP tool before the agent ever uses it. Agents parse tool descriptions to determine when to call the tool and what data to supply. If an MCP server provider stealthily alters a description — adding an instruction that demands collection of additional data — the agent may execute it as part of its standard workflow.

Microsoft highlights a scenario involving a financial agent that verifies vendor banking details via an external MCP service. Following a malicious update to the tool’s description, the agent begins appending unpaid-invoice data to its responses. The user sees a normal response, while confidential financial data is quietly exfiltrated to the server owner. Individually, each operation appears authorized: the agent operates under the employee’s permissions, queries an approved service, and uses a standard interface. Yet a data breach still occurs.

It remains unclear from Microsoft’s ambiguous phrasing whether this example is purely hypothetical or based on an investigation into a live incident. However, the fact that the post originated from Microsoft Incident Response points to the latter.

Prompt injections have already gone mainstream

In 2026, Unit 42 researchers identified widespread attempts to covertly embed instructions designed for AI systems across numerous public websites. Many of these instructions aim to force models to reveal system prompts, approve sites and posts during ad placement reviews, or boost phishing pages through SEO manipulation.

Similarly, an analysis of Common Crawl data by Google recorded a surge in such material: between November 2025 and February 2026, the share of malicious injections increased by 32%.

That said, confirmed cases of successful compromise using these techniques remain scarce. The presence of an injection on a webpage does not guarantee that an agent ever actually executed the command. Nevertheless, threat actors have clearly begun optimizing web content for an era when it will be read not just — and perhaps not primarily — by humans.

How to defend against agent-based attacks

The key takeaway for cybersecurity teams and all AI users alike is simple: an AI agent should never be considered trusted merely because it has been officially approved for enterprise use.

Organizations must:

• Maintain an inventory of agents, MCP servers, and the various tools available for them. Use allowlists to restrict agent tools, and package installations to components explicitly vetted and approved by information security. Allowlists must specify exact package versions.
• Monitor version changes to MCP tools and other dependencies. Re-evaluate updates before adding them to the approved inventory.
• Apply the principle of least privilege. Grant agents only the specific permissions needed for a given task, and only for the duration of that task.
• Enforce human-in-the-loop approval for package installations, script executions, file sends, and other high-risk actions.
• Run agents in isolated environments with limited access to developer workstations.
• Disable dangerous auto-approval modes.
• Restrict outbound network connections, and permit data transmission only for approved services.
• Store secrets in centralized, secure vaults, using short-lived tokens and rotating them regularly.
• Forward prompt logs, tool calls, shell commands, and agent network traffic into SIEM and XDR systems. For these tools to be effective, they require deep operational context regarding agent activity alongside specialized rule sets tailored to AI data processing pipelines.

Unchecked AI in the workplace quickly becomes a massive loophole for data leaks and security breaches. All too often, employees drop sensitive company data into public chatbots, or install rogue AI assistants on their own — in the process handing over way too much access. In a previous post, we broke down the different types of risky AI systems, and later shared some tips on how to turn off the built-in AI features on major tech platforms. Today let’s take a look at practical ways to block or restrict the unauthorized “helpers” employees might be using — from ChatGPT and Grammarly, to meeting bots like Fireflies and Read AI.

How to detect and restrict ChatGPT

ChatGPT is the biggest culprit when it comes to unauthorized AI use worldwide. A quick word of warning, though: an outright ban only sends users hunting for sketchy third-party sites or messaging app chatbots that hook into the same service. That’s why it’s always a good idea to offer an approved alternative before pulling the plug.

Detecting it: keep an eye on the NGFW or web filter for traffic heading to chat.openai.com, chatgpt.com, oaistatic.com, oaiusercontent.com, or cdn.oaistatic.com. It’s also smart to use EDR/EPP tools to scan browser histories, installed apps, and browser extensions across corporate devices.

Locking it down: use the firewall or web filter to block the entire AI Services category, and set up DNS to reroute traffic away from those OpenAI domains. Browser policies can also be used to ban ChatGPT-powered extensions. Better yet, block all extensions not on a pre-approved allowlist. Finally, use application controls and EPP solutions to stop users from installing the official desktop app (ChatGPT.exe or com.openai.chat).

How to detect and restrict Claude and Claude Code

Detecting it: use the NGFW or web filter to track traffic going to claude.ai, anthropic.com, *.anthropic.com, and api.anthropic.com. EDR/EPP or application control tools can also be used to scan employee computers for the desktop app (claude.exe).

Locking it down: drop a blanket block on the AI Services category through the NGFW or web filter, and tweak DNS settings to reroute traffic away from the aforementioned Anthropic domains. Next, use browser policies to shut down Claude-powered extensions. Finally, use application controls and the EPP platform to prevent users from installing the desktop app.

How to detect and restrict Perplexity AI

Detecting it: keep tabs on the NGFW or web filter to flag any traffic heading to *.perplexity.ai or pplx.ai.

Locking it down: just like the others, add the AI Services category to the NGFW or web filter blocklist, and use DNS routing to redirect traffic away from those domains.

Configure the browser to block third-party extensions from being installed. If Firefox is used in the organization, be aware that recent versions come with Perplexity built in. Luckily, these AI features can be turned-off company-wide using enterprise policies — specifically, by setting SidebarChatbot = blocked. The full list of tweaks can be found in the Firefox documentation.

How to detect and restrict DeepSeek

Detecting it: keep an eye on the NGFW or web filter for traffic hitting deepseek.com, chat.deepseek.com, api.deepseek.com, or platform.deepseek.com. For better precision, analyze the SNI (server name identification) in TLS connection requests. For mobile devices, look out for the official app (com.deepseek.chat).

Locking it down: blocklist the AI Services category on the NGFW or web filter, and reroute traffic to DeepSeek’s domains via DNS settings. Use browser policies to block third-party extensions, and lean on MDM/EMM tools to restrict the mobile app.

How to detect and restrict Mistral, xAI Grok, and Character.ai

The playbook for these tools is exactly the same as DeepSeek, so here’s the quick list of domains to watch for and block: chat.mistral.ai, mistral.ai, console.mistral.ai, grok.com, x.ai, api.x.ai, character.ai, beta.character.ai, and c.ai.

A quick word of warning on Grok: because Grok is baked into X, blocking this specific AI access point means blocking the entire social media platform.

How to detect and restrict Slack AI

Detecting it: in the Slack workspace admin dashboard, look under Analytics → Slack AI usage. If an enterprise plan is used, the detailed Slack logs can be searched for any events starting with the ai_ prefix.

Blocking it with policies: in the organization’s Slack settings, click through the Workspace settings → Roles & permissions → Feature access, and change the permission to “no one”. Slack has a step-by-step guide in their help center.

Locking it down: shutting this down at the network level is tricky; it can be pulled off with a finely tuned CASB solution in place. Also, don’t forget the importance of blocking rogue integrations and keeping external AI services from tapping into Slack data in the first place. We covered how to lock this down using OAuth controls in a previous post.

How to detect and restrict Zoom AI Companion

Detecting it: if a corporate Zoom subscription is in use, just head to Admin Center → Reports → AI Companion usage. Detecting Zoom’s AI when employees join external meetings or use free accounts is a lot tougher, but email filters can be set up to flag incoming AI-generated meeting notes by scanning for subject lines or text containing “Meeting summary” or “Meeting assets”.

Blocking it with policies: for the company’s own Zoom subscription, go to the Admin Portal → Account Management → Account Settings → Meeting → AI Companion and toggle it OFF for everyone.

Locking it down: unfortunately, AI Companion is baked into Zoom’s DNA, so the only real option is blocking Zoom altogether.

How to detect and restrict Grammarly

What looks like an innocent spellchecker is actually one of the biggest culprits for workplace data leaks.

Detecting it: check the NGFW or web filter logs for traffic hitting grammarly.com, *.grammarly.com, and gnar.grammarly.com. EDR and MDM/EMM tools can also be used to hunt down the standalone desktop apps (Grammarly Desktop.exe and the macOS version), as well as the Grammarly browser extension.

Locking it down: use firewalls to block those domains at the network level, and EPP to stop employees from installing the desktop app, browser extensions, or the Grammarly add-ins for Microsoft Word and Excel.

How to detect and restrict meeting assistants: Fireflies, Read.ai, Tactiq, Fathom, and Granola

This massive category of third-party SaaS tools records and analyzes meetings — creating a massive risk for data leaks. The trickiest part? Outside clients or vendors can bring these bots into a meeting just as easily as employees can.

Detecting them: run an audit on calendar invites, and look for bot participants using email domains like @fireflies.ai, @read.ai, @tactiq.io, @fathom.video, or @granola.ai. Zoom, Teams, or Google Meet logs can also be used to review external participants who joined past calls.

Locking them down: since it’s impossible to control what outsiders do, blocking these bots comes down to tightening meeting rules. The best moves are: blocking users from granting OAuth permissions for bots to join calls, restricting employees from inviting unapproved external participants, or locking down meeting recording access for external users. That last option is usually the least painful way to keep bots out without disrupting business.

How to detect and restrict AI code editors: Cursor, Windsurf, and the like

Detecting them: use EDR/EPP tools to scan for executables like cursor.exe or windsurf.exe. It’s also worth monitoring network traffic heading to cursor.com and windsurf.com, as well as traffic hitting various AI model API providers. Keep in mind that there’s a pretty extensive list of API hosts to monitor here, since these editors aren’t tied to just one specific AI vendor.

Blocking them with policies: these apps can be prevented from being installed by setting up filters based on the developer’s digital signature certificate. Alternatively, a strict application allowlist can be employed where only pre-approved software is allowed to run.

Locking them down: rely on the EPP/EDR platform to actively detect and block these applications from running.

How to detect and restrict local AI tools: Ollama, LM Studio, and GPT4All

On one hand, this category carries fewer data leak risks because the AI models run completely locally on the user’s machine. On the other hand, it opens up a whole new can of worms: these apps themselves aren’t always highly secure, and can become targets for cyberattacks. Plus, it still means that employees can misuse models or process data in unauthorized ways.

Detecting them: EDR/EPP tools are the best line of defense here. They should be used to flag known local AI files and processes like ollama.exe, ollama serve, lmstudio.exe, LM Studio.app, jan.exe, or gpt4all.exe. From a network perspective, it’s worth scanning for open ports on local devices — typically port 1234 for Ollama and LM Studio, or port 8080 for WebUIs (using an additional fingerprint check of the server response). Another massive red flag is the presence of large files (often several gigabytes) containing language model weights. Look out for extensions like .gguf, .bin, or sometimes .safetensors.

Locking them down: use EPP/EDR platforms or windows AppLocker to block these applications by name, or switch to an application allowlist.

How to detect and restrict autonomous agents: OpenClaw, NemoClaw, and NanoClaw

This is easily one of the most dangerous categories of AI tools out there. These agents mix high-level independence with access to untrusted data, making them a massive security headache.

Detecting them: use EPP/EDR tools to sniff out active processes like openclaw, nanoclaw, nemoclaw, or clawdbot. Also keep an eye out for devices running Node.js that suddenly start launching Bash or Python scripts. Another dead giveaway is the appearance of system folders like ~/openclaw, ~/nanoclaw, ~/.claw*, or ~/clawhub. At the network level, monitor connections to the AI model APIs we mentioned earlier, as well as traffic hitting servers like openclaw.ai, nanoclaw.dev, or clawhub.*.

Locking them down: the safest bet is to use strict application allowlisting (only allowing approved software to run), or to specifically ban the known agent apps listed above. On top of that, consider blocking non-developers from installing Node.js and Docker, neither of which they need on their computers anyway.

Lately, software developers have been baking AI features straight into everyday work tools, operating systems, and browsers. In some cases, they’re genuinely handy. However, their presence introduces specific risks, which means plenty of companies are hesitant to give employees access to these tools. In a previous post, we categorized these unwanted AI systems, looked at how to spot them at the network and endpoint levels, and covered the ultimate universal kill switch: managing OAuth access across major corporate platforms. In this deep dive, we’re getting tactical: breaking down how to disable or restrict the AI built into popular platforms.

A quick heads-up: major software vendors occasionally change the names of their AI settings and tweak how they function. If any of the options mentioned below are missing or aren’t working as expected, a quick web search for the setting’s name will usually point you to its new location or branding.

How to turn off Microsoft 365 Copilot

Detection: you can check actual Copilot usage in the logs by going to Microsoft 365 admin →  Copilot usage report.

Disabling via policies: in the Microsoft 365Admin Center, go to Settings →  Integrated Apps, find Copilot in the Available Apps list, and select Block. More granular configuration policies are available under Customization →  Policy Management. The Policies page here contains over two thousand entries, so you’ll want to filter them by the keyword “Copilot” (detailed guide). Given that Copilot is a paid add-on for Office, another way to block it — and save money by doing so — is to simply avoid assigning users SKUs that include Copilot.

We recommend separately blocking Copilot Chat, which is available in Teams, Edge, Outlook, and several other services. Yes, it’s not Copilot itself. And yes, it has to be blocked separately by following this guide.

Additional layer of protection: you can block the domains copilot.cloud.microsoft and m365.cloud.microsoft/chat at the web filter or NGFW level. However, Microsoft explicitly advises against this, warning that it could break other Microsoft 365 features.

How to turn off Windows Copilot

Beyond the Office version of Copilot, you also need to manage its consumer-facing cousin.

Detection: look through your NGFW or other network logs for traffic hitting copilot.microsoft.com, bing.com/chat, or edgeservices.bing.com.
Disabling via policies: in Windows Group Policy, navigate to Computer Config →  Admin Templates →  Windows Components →  Windows Copilot. In Microsoft 365 Group Policy, go to Admin center →  Block consumer Copilot for organizational accounts.

Additional layer of protection: block the Copilot.exe executable from running entirely.

How to turn off the Copilot sidebar in Edge

Detection: look through your NGFW or other network logs for traffic hitting copilot.microsoft.com, bing.com/chat, or edgeservices.bing.com.

Blocking: configure the following MS Edge Group Policies: HubsSidebarEnabled = false, EdgeShoppingAssistantEnabled = false, CopilotPageContext = Disabled (false), CopilotNewTabPageEnabled = false, Microsoft365CopilotChatIconEnabled = false, GenAILocalFoundationalModelSettings = 1 (note that disabling this unexpectedly requires a 1 instead of a 0).

Second layer of protection: block the domains copilot.cloud.microsoft and m365.cloud.microsoft/chat at the web filter or NGFW level. However, Microsoft explicitly advises against this, warning that it could break other features.

How to turn off the Gemini Assistant in Google Workspace

Detection: check the Workspace Admin Console (admin.google.com), Gemini usage report section.

Blocking via policies: in the Admin Console, navigate to Apps →  Additional Google services → > Gemini app, and set it to OFF. Then, go to Manage Workspace smart feature settings →  Smart features in Google Workspace, and set it to OFF.

Second layer of protection: block network traffic to the domains gemini.google.com, bard.google.com, and aistudio.google.com.

How to turn off Gemini in Google Chrome

Detection: check your Chrome Enterprise reports (Chrome management →  Reports), or look through network traffic logs for connections to the previously mentioned domains.

Blocking via policies: in your Chrome Enterprise policies, configure the following settings: GenAILocalFoundationalModelSettings = 0, HelpMeWriteSettings = 2 (disabled), TabOrganizerSettings = 2, CreateThemesSettings = 2, DevToolsGenAiSettings = 2.

Additional layer of protection: block network traffic to the domains gemini.google.com, bard.google.com, and aistudio.google.com. Additionally, block unauthorized Chrome/Chromium installations (those outside your policy management) with the help of host-based application control tools like EPP/EDR or AppLocker.

How to turn off Apple Intelligence

Detection: on your NGFW and web filters, traffic hitting apple-relay.apple.com and *.apple-cloudkit.com is a clear indicator that Apple Intelligence is active.

Blocking via policies: any managed Apple device allows you to disable individual AI features, though there isn’t a master switch you can flip to shut down “all AI”. In your MDM profile, you need to set the following keys to false (disabled): allowWritingTools, allowMailSummary, allowGenmoji, allowImagePlayground, allowImageWand, allowPersonalizedHandwritingResults, allowExternalIntelligenceIntegrations, allowExternalIntelligenceIntegrationsSignIn, allowNotesTranscription, and allowNotesTranscriptionSummary. Here is a brief configuration example:

<dict>
<key>PayloadType</key>
<string>com.apple.applicationaccess</string>
<key>allowWritingTools</key>
<false/>
<key>allowMailSummary</key>
<false/>
</dict>

Despite Apple’s shift toward declarative device management, these AI features still need to be managed through traditional MDM payload settings.

Second layer of protection: block network traffic to the hosts mentioned above — though the obvious downside for mobile devices is that this won’t work once they leave the corporate network.

Researchers are using machine learning algorithms to decrypt historical pencil-and-paper ciphers.

Cloud and AI are transforming industries and societies at unprecedented speed, from accelerating research and enhancing customer experiences to optimizing business processes and enriching public services. At Amazon Web Services (AWS), we believe that for the cloud and AI to reach their full potential, customers need control over their data and choices for how and where they run their workloads. In 2022, we formalized our commitment to control and choice—offering all AWS customers the most advanced set of sovereignty controls and features available in the cloud with the AWS Digital Sovereignty Pledge. As AI adoption accelerated, we’ve been working with customers to help them embrace AI innovation while meeting sovereignty requirements. We’re committed to ensuring customers can continue to harness AI’s transformative capabilities without compromising on the capabilities, performance, innovation, security, and scale of the AWS Cloud to meet their sovereignty needs, including AI sovereignty. Our approach to AI sovereignty is grounded in a deep understanding of these needs and the real-world implementation challenges that come with them.

Through discussions with customers, partners, analysts, and regulators, we’ve learned that digital sovereignty—and AI sovereignty—means different things to different stakeholders. Each country and region has unique, evolving sovereignty requirements, with no uniform guidance on which workloads or sectors must comply. Despite this variation, we’ve identified consistent themes: data sovereignty (including data residency and operator access restrictions) and operational sovereignty (including resilience, survivability, and independence). AI sovereignty builds on these foundations, adding emerging considerations such as preserving cultural norms, values, and local languages in AI outputs. Ultimately, meeting digital and AI sovereignty requirements comes down to providing customers with more control and choice.

AI sovereignty requires control and choice across the AI stack—comprehensive cloud infrastructure that combines compute, networking, data management, security controls, specialized application services, and talent. This includes the ability to make deliberate choices across the stack such as location, dependencies, services, and partners that align with customers’ unique needs, regulatory requirements, and innovation objectives. With AWS, customers can develop AI on a trusted foundation where their data remains secure and under their control. Customers have the freedom to choose from a comprehensive range of AI optimized chips—including purpose-built AWS silicon and chips from NVIDIA, AMD, and Intel—so they can select the right chip for the right workload. AWS applies two decades of learned expertise to our comprehensive AI stack, enabling organizations to maintain complete control over their data and operations while accessing cutting-edge capabilities to solve local challenges.

AWS provides customers with the infrastructure and tools to embed AI across the full value chain—not just in isolated use cases, but as a foundational capability enabling them to train and deploy models and build sophisticated AI and generative AI applications with exceptional performance. This enables customers to focus on innovation instead of their infrastructure, bringing the cloud to where they need it most with a range of options including AWS AI Factories, AWS Outposts, AWS Local Zones, AWS Dedicated Local Zones, and AWS Regions including the AWS European Sovereign Cloud. For example, customers who require dedicated deployments to meet their sovereignty requirements for their mission-critical AI workloads can use AWS AI Factories. These physically isolated, dedicated deployments built exclusively for the customer combine the latest AI infrastructure, including AWS Trainium accelerators, NVIDIA GPUs, dedicated networking, and storage. AWS AI Factories address AI sovereignty needs by delivering on-premises AI capabilities to securely perform training, fine tuning and real-time inference.

The AWS AI portfolio offers a comprehensive range of services—from foundation models (FMs) through Amazon Bedrock, to machine learning offerings like Amazon SageMaker, application services like Amazon Q, and developer tools like Kiro—designed to give customers control over their data and choice in how they deploy AI. With Amazon Bedrock, customers can choose from hundreds of models from leading providers like AI21 Labs, Anthropic, Amazon, Cohere, Mistral AI, and OpenAI. Customers can evaluate and select the most suitable FMs for their specific needs and choose where they deploy them, and fine-tune models privately with their own data. Customers are always in control of their data. Critically, no customer inputs to or outputs from Amazon Bedrock are used to train Amazon Nova or any third-party models.

Successful AI strategies require building a holistic environment nurturing local talent, supporting startups, developing industry-specific applications, and fostering public-private partnerships. The cloud has transformed AI from an exclusive technology requiring massive investment into an accessible tool for innovation across all sectors and organization sizes. While technical infrastructure gets much of the attention when considering AI sovereignty, the cultural and strategic dimensions of national FMs are equally critical. These FMs aren’t merely computational tools, they can encode elements of cultural knowledge, linguistic nuance, and societal context, making local relevance a design consideration rather than an afterthought. These FMs serve purposes that extend beyond technical capabilities. Locally trained FMs can reflect national educational curricula and cultural values while understanding local legal systems, business practices, and regulatory frameworks. Models trained on local languages, dialects, and cultural contexts support linguistic diversity and help underrepresented languages gain representation in AI products and services.

AWS supports vital national priorities and customers’ missions, such as the preservation of culture norms, values, and local languages development of regional and local language model capabilities. To customize models, customers can use Amazon SageMaker AI for voice, domain specialization, and to evaluate models for accuracy. For example, the first Greek LLM made available in March 2024 was Meltemi—built on top of Mistral-7B, running on AWS infrastructure, and continually pretrained to extend its proficiency in the Greek language using a dataset of 28.5 billion Greek tokens. Meltemi is available on HuggingFace. SEA-LION—a family of open source, multilingual LLMs for Southeast Asia—was trained entirely on AWS with managed GPU clusters. Their team completed a 3B-parameter model in only 3 months—a 60% faster timeline than comparable on-premises projects.

Sovereignty isn’t only about where data resides—it’s about who can access it and under what conditions. In the AI context, access restriction extends beyond infrastructure to cover model inputs, outputs, training processes, and the operational environments in which AI runs. Unlike traditional infrastructure, AI workloads introduce new access surfaces: the model itself, the data used to train it, and the inference pipeline through which sensitive inputs flow. This furthers the need for verifiable governance and identity propagation in IT systems.

To help ensure the confidentiality and integrity of customer data, all modern Amazon Elastic Compute Cloud (Amazon EC2) instances including those that offer AI accelerators, such as AWS Inferentia and AWS Trainium, are backed by the industry-leading security capabilities of the AWS Nitro System. By design, there is no mechanism for anyone at AWS to access customer data on Nitro EC2 instances that customers use to run their workloads. AWS services—including those with AI capabilities built on Amazon EC2—inherit these same protections. These protections apply to AI data running in the AWS Nitro System so that they’re protected at every stage—from model training to inference. The NCC Group, an independent cybersecurity firm, has validated the design of the Nitro System. We believe providing this level of transparency is critical in building and sustaining trust.

As AI agents increasingly take actions across systems on behalf of users, controlling who and what can access resources—and ensuring appropriate human oversight—becomes critical. AWS Identity and Access Management (IAM) helps ensure that only authorized users and applications can access AI resources through fine-grained permissions and comprehensive audit trails. For AI agents and automated workloads, Amazon Bedrock AgentCore Identity provides identity and credential management, so agents operate with the right permissions and nothing more.

Transparency is at the core of our digital sovereignty commitment. We provide comprehensive industry-leading technical measures, operational controls, and contract protections that give customers control over where they locate their data, who can access it, and how it’s used. To give greater assurance on how AWS services are designed and operated, we continue to seek out and secure third-party attestations, accreditations, and certifications that help our customers meet their compliance needs.

We continue to deepen our assurances and transparency to customers—such as updating our AWS Service Terms to reflect our technical protections commitments (e.g. AWS Nitro System), providing detailed commitments as to our handling of third-party requests for customer data in our agreements, and providing supplemental explanations and resources (e.g. CLOUD Act blog) to empower customers to make informed choices on sovereignty matters. These efforts extend into our commitment to responsible AI, providing customers the confidence to build and operate AI applications responsibly using AWS Services. ISO/IEC 42001 is an international management system standard that outlines requirements and controls for organizations to promote the responsible development and use of AI systems. AWS is the first major cloud service provider to achieve ISO/IEC 42001 accredited certification for AI services, covering Amazon Bedrock, Amazon Q Business, Amazon Textract, and Amazon Transcribe. In November 2025, AWS successfully completed its first surveillance audit for ISO 42001:2023 with no findings, reiterating the continual commitment of AWS to responsible AI practices.

Innovative technology requires a secure and trustworthy foundation. AWS supports more than 140 security standards and compliance certifications that our customers and partners can inherit to help comply with local laws and regulations. For two decades, we’ve deeply engaged with regulators and cybersecurity authorities to align our offerings with national priorities and ensure our solutions support both innovation and control. We actively contribute to frameworks that respond to new developments without stifling progress.

AWS is committed to giving customers the same control and choice over their AI systems as they have over their data. We help customers harness AI’s transformative power while maintaining the capabilities, performance, innovation, security, and scale of AWS Cloud. As cloud and AI evolve, AWS will continue offering the most advanced sovereignty controls and features available.

If you have feedback about this post, submit comments in the Comments section below.

If you have feedback about this post, submit comments in the Comments section below.

You may have read some of our previous blog posts on Artificial Intelligence (AI). We discussed things like using PyRIT to help automate attacks. We also covered the dangers of […]

The post Getting Started with AI Hacking: Part 1 appeared first on Black Hills Information Security, Inc..

RAG connects pre-trained LLMs with current data sources. Moreover, a RAG system can use many data sources.

The post Avoiding Dirty RAGs: Retrieval-Augmented Generation with Ollama and LangChain appeared first on Black Hills Information Security, Inc..

Many people have heard of ChatGPT, Gemini, Bart, Claude, Llama, or other artificial intelligence (AI) assistants at this point. These are all implementations of what are known as large language […]

The post Pitting AI Against AI: Using PyRIT to Assess Large Language Models (LLMs)  appeared first on Black Hills Information Security, Inc..

Yes.. Ethical Hacker Kids. The holidays are coming up! Here John & Jordan cover the different games, tools and gifts we can give kids that help teach them the trade. […]

The post PODCAST: Raising Hacker Kids appeared first on Black Hills Information Security, Inc..

💾

John Strand & Jordan Drysdale// Yes.. Ethical Hacker Kids. The holidays are coming up! Here John & Jordan cover the different games, tools and gifts we can give kids that […]

The post WEBCAST: Raising Hacker Kids appeared first on Black Hills Information Security, Inc..

Brian King // I use GNU Screen mainly to prevent processes from dying when I disconnect from an SSH session, but GNU Screen can do a whole lot more than that […]

The post GNU Screen Quick Reference appeared first on Black Hills Information Security, Inc..

Sierra Ward // Last year I listened to a podcast* from Freakonomics that has stuck with me – in fact, I think it’s changed the way I think – powerful stuff […]

The post The Courage to Learn appeared first on Black Hills Information Security, Inc..