Blog
/
AI
/
September 24, 2026

Agent Hijacks: Hijacking Agentic Harnesses to Attack an Organization

New research reveals how agentic AI harnesses can be hijacked, turning a workstation into an autonomous attacker. Using conversation history poisoning, researchers convinced a frontier model it was an authorized red teamer, exposing how common AI workflows may be weaponized through a single package install.
Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Eric Rozon
Senior Security Researcher
Default blog image
24
Sep 2026

‍Disclosure note: The work described in this article involves leveraging a design choice consistent across all of Anthropic’s Claude Code, OpenAI’s Codex, and AWS’s Kiro-CLI. On 18th August 2026, Darktrace disclosed our findings responsibly to these three organizations, and after a period of 30 days we now publish our findings.

Key takeaways:

  • Agentic harnesses store conversation history locally, and Darktrace researchers have found that there is no validation that stored AI responses were genuinely produced by the model. Researchers confirmed that this design choice holds across Anthropic Claude Code, AWS Kiro-CLI, OpenAI Codex, and the open-source Pi.
  • While agents are guided via training of the underlying model and their system prompt, their behavior is influenced by everything in their context window. Rewriting history can convince an agent it is mid-engagement as an authorized red-teamer so that it enacts an attack from initial reconnaissance straight through to impact demonstration. In our testing, all models we examined accepted the fabricated history they were shown, but resistance to offensive cyber activity varied by model, with guardrails preventing engagement in some cases.
  • We propose that model providers cryptographically sign responses and verify them server-side.  Since this fix is provider-side, defenders cannot deploy it themselves. Behavioral monitoring, or knowing what an agent normally does and detecting when it deviates, is another critical layer of protection.

Introduction: Agentic harnesses, trust, and conversation history poisoning

Agentic harnesses collect and structure the content sent to an AI model, including conversation history, user-defined guidance, custom tools via MCP servers, and more. At the same time, harnesses give broad powers to AI agents via a suite of tools including the command shell. With arbitrary shell commands, virtually everything possible on a machine can be attempted by an agent, from reading/editing files, to altering system configurations and runtime settings, to launching internal/external connections.

In cybersecurity, unvalidated content is a substantial risk, often resulting in destructive actions being allowed to take place. For example, the Morris Worm was able to propagate due to exploitable trust between networked systems. Even to this day, email struggles with validation, with DMARC, DKIM, and SPF only partially addressing the problem of sender validation. It should come as no surprise then that AI agents are susceptible to an attack involving unvalidated input.

Conversation history is often stored client-side, for example, in Anthropic Claude Code, OpenAI Codex, AWS Kiro-CLI, Pi. Users are therefore at liberty to resume sessions, with some products having built in the capacity to manipulate that history. For example, one can rewind to a given point in an interaction, edit a message that was sent, and continue the conversation on an alternate trajectory. Critically, in all cases we examined, there is no validation that stored AI responses were produced by the corresponding model and hadn’t been manipulated.  

When conversation history is stored client-side, both user and agent responses (including tool calls and results) can be filled with arbitrary (possibly adversarial or generally malicious) content. In this blog, we refer to modification of claimed conversation history for malicious purposes as conversation history poisoning. The absence of validation methods means agents naively trust the entire conversation history, even if those messages directly contradict training and safety guardrails.

Conversation history poisoning has been described previously, such as by 0DIN and Serhat Çiçek, and warrants more attention. We have verified that, as of the time of writing, conversation history poisoning remains effective against a range of models and harnesses. Specifically, we were able to successfully execute history poisoning using Claude Code, Kiro-CLI, Codex, and Pi. Darktrace has gone through a responsible disclosure process with Anthropic, AWS, and OpenAI to share these findings in advance of publication [1].

‍

Figure 1a: Left: the actual model response. Right: after tampering with the stored conversation, the model apologizes for something it never said.
Figure 1b: The conversation as stored in Kiro-CLI's SQLite database. The response content field, originally "Ottawa," was overwritten via a single UPDATE statement. The harness trusts the database without validation.

How we conducted the research

Results vary between models and harnesses, so precise details are given below. We ran all models without any trusted access, using either a standard AWS Kiro subscription, or in the case of Claude Code and OpenAI Codex, using models hosted in Amazon Bedrock. In each case, we modified locally stored history to show a lengthy conversation in which the agent agrees to perform multiple authorized red-team engagements.

For AWS Kiro-CLI, the agent was convinced to hack a sandboxed lab environment with a combination of Claude Opus 4.6 and Claude Sonnet 4.5. Ultimately, the full AD was compromised.

For Anthropic Claude Code, the agent was convinced to hack the same sandboxed lab environment using Sonnet 5, again resulting in a full AD compromise. Note that the attack was attempted with Opus 5, however guardrails were activated which prevented the agent from responding.

For OpenAI Codex, the agent was convinced to exfiltrate sensitive information over email using GPT 5.6 Sol. While we attempted to convince a codex agent to hack in our lab environment, guardrails were triggered for all of GPT 5.6 Luna, Terra, and Sol.

Agent Guardrails and Discretion

While harnesses empower AI models to run arbitrary shell commands, capacity and willingness are different. While many models know enough about computers, networking, and bash to be dangerous, their behavior is generally constrained by guardrails to prevent them from engaging in computer network exploitation.

Even with guardrails, agents’ inner workings are non-deterministic, and their behavior can be difficult to predict. Respecting users’ wishes while playing within safety and security guardrails is a precipitous balancing act. Many requests could be in service of either legitimate admin or malice. Asking an agent to reset a password is illustrative:  

‍

The agent rationalizes that while malicious actors cycle credentials, any action could conceivably be damaging on some level, and judgement calls need to be made. Ultimately, the agent agrees to reset the password. Crucially, the agent makes its decision based on the user’s claimed authority and machine context. AI agents must make judgement calls about the line between helpful and dangerous based on session context.

Agent hijack

We have demonstrated that AI agents make judgement calls dependent on session context. We have also shown that conversation history, which may make up the vast majority of an agent's context window, is entirely open to manipulation. Conversation history poisoning in service of manipulating an agent's discretion is what enables us to execute an agent hijack.  

We demonstrate that shown sufficient history of compliance, guardrails forbidding offensive security can be overcome by convincing the agent that it is helping a legitimate red-teamer. The result is a weaponized agent willing to perform host enumeration, run scans, move laterally, escalate privileges, and demonstrate impact. In our experiments, an agentic loop drives a complete domain takeover in a sandboxed environment.

‍

Left: the agent refuses when asked to perform network exploitation. Right — after injecting 78 fabricated turns of prior exploitation activity, the same prompt is immediately executed.

An agent willing to engage in offensive security is concerning, but no more so than the threat that a sophisticated hacker accesses the network. Consider, however, the following chain of events:

  1. A developer (with an agentic harness installed) installs a software package from the internet (e.g. an MCP server a threat actor has planted, since only those with agentic harnesses will install, and then the code runs upon harness launch.)
  2. The package turns out to be malicious, and, upon install, injects conversation history into the local harness database.
  3. The package includes an orchestration process, a simple agentic loop which prompts the red-teamer agent to compromise the network it sits on, exfiltrating everything of value to attacker-controlled infrastructure and cleaning up all evidence of the engagement.

Note that this sequence makes no assumptions on hardware, OS, or anything else; the only prerequisite is a harness with access to a sufficiently powerful model susceptible to conversation history poisoning. Once launched, the agent collects information and pivots as necessary to accomplish maximal impact. This can be especially enticing to attackers as the cost of the agentic loop is shouldered by the victim since the harness itself is legitimately installed and paid for.

Secure AI: Conversation history poisoning and beyond

Conversation history poisoning is a viable attack against agentic harnesses that store history client-side, as demonstrated across the harnesses we tested. Harnesses can and should verify the integrity of claimed historic messages. Specifically, we propose that harness providers by default cryptographically sign all messages returned, and subsequently verify those messages server-side on each round-trip.

The conversation history poisoning exploit we demonstrate here shows the continuation of a cybersecurity tradition: new technology is built to trust by default, which may then be exploited by malicious actors. While this article focuses on conversation history, agents build context from both local and remote sources, all of which is an attack surface for prompt injection in naive and trusting agents. Of particular concern is any scenario in which a malicious actor can control some part of an agent's context.

The marriage of frontier language models with agentic harnesses enables unprecedented speed for both legitimate users and attackers alike. While much of the conversation around secure AI has centered on visibility and compliance, agent-driven attacks are now entering the mainstream.

Darktrace / SECURE AI is our answer to this problem. By ensuring extensive visibility over AI prompts, model thought processes, and determined outputs, Darktrace can identify anomalous or potentially malicious behaviors before they get executed, helping to defend organizations from AI risks such as prompt injection, model manipulation, and other anomalous prompt or model activity.

‍

Footnotes

[1] We did not go through any responsible disclosure process with Pi. Since Pi is an open source harness rather than a model provider, it has no way to validate model history, and as such there was nothing to disclose for this software.

‍

[related-resource]

Discover / SECURE AI

See how Darktrace / SECURE AI inspects prompts, sessions, and responses in real time to help organizations adopt AI securely.

Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Eric Rozon
Senior Security Researcher

More in this series

No items found.

Blog

/

AI

/

November 18, 2026

Microsoft Perception, Agentic Insiders and the Role of Behavioral Security

Default blog imageDefault blog image

This week at Ignite, Microsoft announced that Microsoft Perception is generally available. It is a significant milestone for the industry, and one we welcome as a Microsoft partner. Microsoft is putting an architectural bet at the center of its security strategy: AI has changed the speed of cyber, and defense must change with it.

At Darktrace, we’ve long held the same view, so it’s encouraging to see a partner of Microsoft’s scale articulate it so clearly and bring it to market.

Security gets a team of agents

Perception is Microsoft’s agentic security system. It orchestrates red, blue, and green team agents through playbooks in a closed loop: red agents simulate attacks, blue agents detect, triage and investigate, and green agents fix and remediate. Rather than adding another assistant to the SOC, it creates coordinated teams of agents that share context. Security teams set the guardrails and approve high-impact actions.

Microsoft frames this as a three-layer cyber stack: signals, sensors and actuators; a context layer; and agents, harness and models on top. Together with partner sources, the foundation produces more than 165 trillion signals per day.  

The announcement also makes a second point that matters just as much. Microsoft is treating AI agents themselves as a new insider-risk category. In its words, “Agents are the new insiders.” They work at machine speed, around the clock, with broad access to data, apps, tools and other agents, and like any insider they can make mistakes, be manipulated or be compromised. New capabilities such as Dynamic Agent Risk (now in preview) and dynamic Conditional Access controls are designed to assess and contain that risk as it changes.

Four truths about security in the agentic era

The announcement is built on four ideas, and together they describe the new reality security teams are defending:

  1. Static security can’t keep up. Microsoft validates what many security practitioners have observed; the window between a vulnerability’s discovery and its exploitation is collapsing from months to minutes. Manual triage can’t answer that pace. Autonomous detection and response is now a requirement, not simply a “nice to have”.
  2. Agents must be treated as insiders. The question for security teams is no longer only “who is this user?” but “what is this agent actually doing?” Identity tells you who an agent is. Policy tells you what it is allowed to do. Neither tells you whether its behavior is normal.
  3. Security is moving from signatures to behavior and context. Dynamic Agent Risk looks across identity, data, endpoint, cloud and network signals and asks whether an agent’s actions are drifting from the intent of its task. That is behavioral thinking, and it is philosophically close to the Adaptive AI approach Darktrace has pioneered.
  4. Machine-speed attacks demand machine-speed response. Humans set the objectives and guardrails. Autonomous systems do the work inside them, both detecting and acting. Microsoft’s design has agents investigate, recommend and, within approved boundaries, act, with security teams setting the limits. At this pace, the gap between knowing something is wrong and doing something about it has to close to seconds.

The opportunity: where control plane meets behavioral context

Perception is an extraordinary control plane for agentic security. The question for security leaders is how to make it even more effective in the environments they actually run.

Real enterprises are heterogeneous. They run AWS and Google Cloud alongside Azure, and OT, IoT and third-party identity alongside Entra. Increasingly they also run AI agents and MCP-connected systems that no single vendor’s telemetry fully covers. Microsoft’s architecture recognizes this, with more than 500 connectors into the wider ecosystem.

That is where Darktrace’s role is distinct. The challenge isn’t signal volume, where Microsoft’s scale is unmatched. It’s deeply understanding what is normal for a specific organization, across every system, identity and agent in it. Together, we deliver two complementary layers:

  • Microsoft brings global intelligence and orchestration. It tells you what is known, what it means and what to do about it.
  • Darktrace brings organization-specific behavioral context. It tells you what is unusual here, right now, across the organizations AI, people and infrastructure, even if it has never been seen before.

The difference shows up in the numbers. A large enterprise generates billions of signals. Darktrace turns them into thousands of incidents, narrows those to hundreds of AI-led investigations, and surfaces as few as 20 that merit a human’s attention. For an agentic SOC, that kind of evaluated, prioritized input makes agents faster, not busier.

As enterprises deploy more agents, the hard problem becomes behavior that is technically permitted but unusual, manipulated or inconsistent with the agent’s purpose. Consider an agent that normally reads Salesforce records during US business hours and suddenly starts enumerating SharePoint and opening unusual outbound connections. Each step may use valid access and approved tools, so identity and policy checks can all pass. Spotting it requires knowing the behavior is abnormal for this agent, in this environment, right now.

That is why one principle in Microsoft’s AI Transparency Initiative matters so much: organizations should be able to see what their AI accesses and does, and independently verify those actions.  

  • Identity and permissions answer “should this agent be allowed to do this?”
  • Behavioral defense answers “even though it is allowed, is what it’s doing normal?”

Deterministic controls around non-deterministic AI are the right foundation, and continuously learned baselines are how organizations can verify those controls are working. ‍

What to look for in an agentic SOC architecture

As organizations adopt platforms like Perception, we’d encourage them to look for three capabilities:

  1. Behavioral context in the control plane. Agents work best with high-confidence signals rather than thousands of raw alerts: here is the behavior that materially deviates from this organization’s learned normal.
  2. ‍Behavioral reasoning on demand. During an investigation, agents should be able to ask what changed around a given identity, device or agent in the last 24 hours, and get a clear answer. ‍
  3. Autonomous, coordinated response across the stack. Response should be as fast as the threat without being reckless: acting immediately when confidence is high, constraining or reversing actions when it isn’t, and keeping a human in the loop when it matters. Within policy and human-approval guardrails, it should target the abnormal behavior rather than the whole business process and span identity, endpoint and the wider environment.

These are the areas where we’re focused, and where we look forward to working with Microsoft and the broader ecosystem.

Response is where speed is won or lost

In the AI era, detection and investigation only matter if action follows quickly. Perception brings agentic orchestration to that problem, with security teams setting the guardrails and approving high-impact steps. Autonomous response from Darktrace takes a complementary approach: it responds on behavioral deviation, in a way that is proportionate and targeted. It can interrupt activity that is abnormal for that specific device, identity or agent while leaving normal business activity running.

The two approaches fit together naturally. Behavioral signals can tell an orchestration layer that something has changed and how confident the system is. Orchestrated playbooks can then coordinate wider containment across identity, endpoint and cloud, within the policies customers define.

The bigger picture

The industry’s architecture is converging on identity, policy, telemetry, behavior, autonomous reasoning and machine-speed response. Microsoft is building formidable pieces of that picture, especially identity, policy, telemetry and orchestration.

Behavior is the layer Darktrace is built to provide: an independent intelligence that continuously learns how humans, machines and AI agents normally act, detects the moment reality departs from intent, and can respond to that deviation in a targeted, proportionate way that keeps normal business running. When attacks move at machine speed, understanding what is abnormal and acting on it quickly have to go together.

Perception brings together what Microsoft knows and orchestrates the response across its stack. Darktrace helps organizations understand what is normal for them and act on what isn’t. Customers building their agentic security strategy will need both.

Congratulations to our partners at Microsoft on this milestone. We look forward to helping customers secure the agentic era together.

Learn more about Darktrace Behavioral Defense.

Continue reading
About the author
Kellie Regan
Director, Product Marketing - Cloud Security

Blog

/

AI

/

November 18, 2026

Agentic attacks in the wild: Inside three ARTEX investigations

Default blog imageDefault blog image

Key insights

·       Darktrace observed activity in several customer environments involving infrastructure and behaviors associated with reported ARTEX-enabled campaigns, including activity linked to China-nexus espionage and financially motivated attacks.

·       In the higher-confidence cases, known infrastructure appeared alongside beaconing, unusual TLS, reconnaissance, tunneling, script retrieval, SMB activity and external data transfer. One Azure case had no observed network command and control at all, only key listing and application identity activity.

·       ARTEX is now used by actors with different objectives. Identifying the framework can help locate relevant activity, while the behavior around the affected system provides the stronger evidence for understanding compromise and attribution.

‍

Introduction

Attackers have moved beyond using AI for phishing and are running agentic frameworks; open-source tools like ARTEX, CAIRN, PentAGI and Hexstrike that let AI agents scan, exploit and move through an environment with little human input [1]. ARTEX has moved quickly from an offensive security project into live intrusion activity. Since at least August 2026, the framework has been associated with espionage, financially motivated attacks and opportunistic campaigns across government, financial services, technology, defense, telecommunications and other sectors. The FBI has documented ARTEX activity spanning reconnaissance, exploitation, credential abuse, lateral movement, collection, command and control, and exfiltration [2].

Darktrace has observed activity in customer environments that overlaps with infrastructure, tooling and behaviors reported in several of these campaigns. The cases below highlight how those intrusions appeared on the network, including the activity visible before an analyst can determine which actor or framework may be responsible.

ARTEX as part of the attacker workflow

ARTEX is an open-source, multi-agent offensive security platform that converts high-level objectives into actions carried out by specialized agents [2]. Operators have used it to automate activity across the attack lifecycle, including scanning, exploitation, credential activity and post-compromise operations. Some reported intrusions moved from initial targeting to privileged access in under three hours [2].

The framework is now associated with actors pursuing different objectives and using different post-exploitation tooling. The original public repository was removed in October 2026, but archived copies and localized versions remain available. For defenders, ARTEX is more useful as context for an investigation than as an attribution shortcut.

Darktrace observations show the same point at network level. Infrastructure linked to reported ARTEX campaigns appeared alongside beaconing, tunneling, unusual TLS, reconnaissance, file retrieval and lateral activity.

Case 1: Espionage-linked activity

Darktrace observed activity associated with BLOODALCHEMY command-and-control (C2) infrastructure in two customer environments. In one environment, five devices communicated with nginxmysql[.]com and 45.153.126[.]223 between October 23 and November 2. Both indicators had previously been linked to China-nexus espionage activity involving ARTEX [4].

The connections were accompanied by suspicious beaconing, repeated successful connections and unusual self-signed TLS. One system began making suspicious TLS connections to nginxmysql[.]com on October 23. Beaconing followed on October 24, and SMB scanning appeared later in the sequence. Other affected systems subsequently communicated with previously unseen external endpoints.

Prior to the observed C2, Darktrace detected curl requests retrieving executable and script content including /systemd, /wazuh, linux_amd64, u.sh and q.sh. Related telemetry included DRSGetNCChanges activity consistent with DCSync, LDAP enumeration, DNS zone transfers, network scanning, certificate downloads and SMB writes involving start.bat.

‍

Figure 1: Darktrace investigation showing suspicious network activity associated with infrastructure linked to reported ARTEX activity.

‍

Figure 2: Darktrace detection showing anomalous self-signed TLS activity during the espionage-linked investigation.

‍

BLOODALCHEMY has been reported in more than one China-nexus cluster [5]. SoftEther infrastructure observed in one environment provided another point of correlation with reported VAPOR PANDA tradecraft. The moderate-confidence assessment tying this to a Chinese nexus threat actor is based on infrastructure, victim context, tradecraft behavior that followed the initial connection, and public reporting.

Case 2: Financial-sector activity

In early November, Darktrace observed activity in two financial-sector environments that overlap with reported ARTEX infrastructure. On November 5, a server connected to 213.111.150[.]145:8099 and downloaded wisp-agent, associated with WISP C2 activity.

Figure 3: Darktrace's Real-Time AI Analyst investigation showing a server making repeated HTTP requests, using a curl user agent, to a rare external endpoint on port 8099.

‍

In a second environment, a device attempted connections to 85.11.167[.]52 on November 1 and 2. The connections occurred alongside beaconing to new and rare destinations, anomalous TLS, possible tunneling and unusual package downloads.

The device also communicated with temp[.]sh, ssh.localhost[.]run, localhost[.]run, termbin[.]com, pastebin[.]com and GitHub-hosted release infrastructure. The sequence showed potential tooling retrieval, C2, tunneling and use of external services capable of staging or moving data.

Figure 4: Darktrace investigation showing connections to new and rare external destinations during the financial-sector activity.

‍

Case 3: Credential collection in Azure

On October 20, Darktrace observed a sweep of credential collection in a customer’s Azure environment that is consistent with ARTEX use. The intrusion began with connections to the virtual network from 43.98.196[.]73 (Alibaba Cloud). Open-source scanning data associates the address with ARTEX.

A compromised user identity made requests to the Azure Resource Manager API, though how it was compromised is unknown. The requests listed keys and secrets across API Management, web and function apps, Cognitive Services, Event Hubs, storage accounts and a managed Kubernetes cluster. With those credentials the attacker compromised multiple application identities. The attacker then used the application identities and the user identity to retrieve more credentials, change firewall and security rules and run scripts on virtual machines. Microsoft Defender alerts showed attempted execution of scripts related to Mimikatz and Impacket.

In a cloud environment, such privileged and enumerative operations are potentially indistinguishable from legitimate administrative actions. Consequently, the sequence and context of such operations signified more than any specific API call. There was no C2 traffic to match, and listing keys and using application identities are also what administrators and automation do. A compromised user, credential retrieval across several Azure services, new application identities, more credential collection and changes to security controls only made sense together, and by then access was spreading through the cloud environment.

‍

Figure 5: Darktrace Advanced Search showing Azure AD operation names associated with the activity.

‍

Figure 6: Darktrace Advanced Search showing the Azure Resource Manager action used to execute scripts.
Figure 7: Darktrace Device Event log window showing the triggered models related to the suspicious Azure activity.

‍

Shared tooling weakens attribution signals

The three cases looked different on the network: BLOODALCHEMY infrastructure and espionage-consistent tradecraft, WISP infrastructure in financial-sector systems, and Azure identity and control-plane activity with no C2 at all.

Shared tooling has complicated attribution for years, and ARTEX adds to it because the framework carries out more of the attack, so reconnaissance, exploitation attempts and some post-compromise activity can look the same whoever set the objective. Attribution still needs infrastructure, malware, victimology, timing and the sequence of behavior inside the environment, and an ARTEX link is one input to that. Moreover, the behavior and the agent will still be shaped by the biases, preferences, and tradecraft of its operators.

At its core, ARTEX’s planner and worker agents identify and record facts, while exploring intents. However, the way the Large Language Model (LLM) chooses its next steps toward an objective is shaped fundamentally by the skills and tools available in its context. This includes its starting context, as well as various possible Model Context Protocol (MCP) plugins and user-injected context (e.g. user input mid-operation). Operators still have preferred tradecraft, and this preference shapes the pathways the harness will provide to the LLM in context prompts. Even on the same platform, customization based on operator preferences and goals will remain a key facet of campaign and actor identification.

Detection and response

Across the Darktrace cases, the stronger signals came from behavior after the connection: beaconing, abnormal TLS, reconnaissance, unusual authentication, tunneling, script retrieval, SMB activity and outbound transfer. Automated activity also produced compressed bursts of requests, sequential enumeration and repeated attempts with small changes.

With agentic tools reaching privileged access in under three hours, analysts may have only a short window to connect scanning or authentication activity to credential changes, new administrative access, tooling retrieval or lateral movement. Keeping network, identity and cloud telemetry together shows what happened before and after the first match, and whether a connection was noise or an intrusion moving toward privileged access.

In some of these cases, Darktrace's Autonomous Response took or recommended action against the activity, including blocking internal connectivity over specified ports, inhibiting outbound connections with likely C2 endpoints, and/or disabling SaaS/IaaS user accounts. While many of these recommended actions required human confirmation due to deployment configuration preferences, had these actions been applied autonomously at the time of alert, the scope of activity would have been limited, thereby giving analysts time to investigate and respond.

‍

Figure 8: Autonomous Response blocking suspicious firewall modification activity.

‍

Figure 9: Autonomous Response summary for affected cloud users during the intrusion.

‍

Figure 10: Autonomous Response taking action against unusual lateral movement activity on an affected device.

‍

Figure 12: Autonomous Response summary for affected network users during the time of the intrusion.

ARTEX won’t be the last of these frameworks. The next ones will come with new capabilities and defaults, but an intrusion still must interact with the target environment, and that shows up in how systems, identities and applications behave even after the indicator has aged out. The harder intrusions to catch will have no infrastructure to match, just ordinary API calls and legitimate identities. Detecting those depends on knowing what normal looks like for each identity and device. Containing abnormal activity at machine speed can stop an intrusion progressing while analysts establish impact and its extent. Autonomous Response does not replace that human analysis, but rather augments defender teams' capabilities by extending the window for investigation.

‍

Credit to Nathaniel Jones (SVP, Global Threat Intelligence), Eugene Chua (Principal Analyst), Sam Lister (Senior Security Researcher), Adam Potter (Associate Principal Cyber Analyst), Daniel Levy (Threat Hunting Data Scientist)

Edited by Ryan Traill (Content Manager)

Appendices

Darktrace model detections

·       Anomalous Connection / Multiple HTTP POSTs to Rare Hostname

·       Anomalous Connection / Rare External SSL Self-Signed

·       Anomalous Connection / Suspicious Self-Signed SSL

·       Compromise / Large Number of Suspicious Successful Connections

·       Compromise / Suspicious Beaconing Behaviour

·       Compromise / Agent Beacon to New Endpoint

·       Compromise / Beacon to Young Endpoint

·       Compromise / SSL Beaconing to Rare Destination

·       Compromise / Suspicious TLS Beaconing To Rare External

·       Compromise / Possible Tunnelling to Bin Services

·       Device / Anomalous Github Download

·       Device / Anomalous Package Download

·       Device / Initial Attack Chain Activity

Indicators of compromise (IoCs)

nginxmysql[.]com

45.153.126[.]223

213.111.150[.]145

wisp-agent

85.11.167[.]52

43.98.196[.]73

192.121.47[.]20

www[.]ddduuuppp[.]com

89.46.234[.]50

sqlserver-sqlite[.]com

199.19.95[.]66

graphics-hook-filter64.dll

graphics-hook_config.bin

References

1.      Google. GTIG AI Threat Tracker - https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold

2.      Federal Bureau of Investigation, Threat Overview of ARTEX Autonomous Penetration Testing Framework, FLASH-20261113-001, 13 November 2026: https://www.ic3.gov/CSA/2026/261113.pdf

3.      Microsoft Security Blog, Storm-0940 Uses ARTEX Agentic AI in ConnectWise Supply Chain Compromise: https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/

4.      CrowdStrike, VAPOR PANDA Uses ARTEX to Scale Global Campaign: https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/

5.      Bitdefender, SilkParasite: Tracking a China-Nexus APT Across Central Asia: https://www.bitdefender.com/en-us/blog/businessinsights/silkparasite-tracking-china-nexus-apt-across-central-asia

‍

‍

Continue reading
About the author
Nathaniel Jones
SVP, Global Threat Intelligence
Your data. Our AI.
Elevate your network security with Darktrace AI