Blog
/
Network
/
November 7, 2022

[Part 1] Analysis of a Raccoon Stealer v1 Infection

Darktrace’s SOC team observed a fast-paced compromise involving Raccoon Stealer v1. See which steps the Raccoon Stealer v1 took to extract company data!
Written by
Mark Turner
SOC Shift Supervisor
Written by
Sam Lister
Specialist Security Researcher
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
Mark Turner
SOC Shift Supervisor
Written by
Sam Lister
Specialist Security Researcher
Default blog image
07
Nov 2022

Introduction

Towards the end of March 2022, the operators of Raccoon Stealer announced the closure of the Raccoon Stealer project [1]. In May 2022, Raccoon Stealer v2 was unleashed onto the world, with huge numbers of cases being detected across Darktrace’s client base. In this series of blog posts, we will follow the development of Raccoon Stealer between March and September 2022. We will first shed light on how Raccoon Stealer functioned before its demise, by providing details of a Raccoon Stealer v1 infection which Darktrace’s SOC saw within a client network on the 18th March 2022. In the follow-up post, we will provide details about the surge in Raccoon Stealer v2 cases that Darktrace’s SOC has observed since May 2022.  

What is Raccoon Stealer?

The misuse of stolen account credentials is a primary method used by threat actors to gain initial access to target environments [2]. Threat actors have several means available to them for obtaining account credentials. They may, for example, distribute phishing emails which trick their recipients into divulging account credentials. Alternatively, however, they may install information-stealing malware (i.e, info-stealers) onto users’ devices. The results of credential theft can be devastating. Threat actors may use the credentials to gain access to an organization’s SaaS environment, or they may use them to drain users’ online bank accounts or cryptocurrency wallets. 

Raccoon Stealer is a Malware-as-a-Service (MaaS) info-stealer first publicized in April 2019 on Russian-speaking hacking forums. 

Figure 1: One of the first known mentions of Raccoon Stealer on a Russian-speaking hacking forum named ‘Hack Forums’ on the 13th April 2019

The team of individuals behind Raccoon Stealer provide a variety of services to their customers (known as ‘affiliates’), including access to the info-stealer, an easy-to-use automated backend panel, hosting infrastructure, and 24/7 customer support [3]. 

Once Raccoon Stealer affiliates gain access to the info-stealer, it is up to them to decide how to distribute it. Since 2019, affiliates have been observed distributing the info-stealer via a variety of methods, such as exploit kits, phishing emails, and fake cracked software websites [3]/[4]. Once affiliates succeed in installing Raccoon Stealer onto target systems, the info-stealer will typically seek to obtain sensitive information saved in browsers and cryptocurrency wallets. The info-stealer will then exfiltrate the stolen data to a Command and Control (C2) server. The affiliate can then use the stolen data to conduct harmful follow-up activities. 

Towards the end of March 2022, the team behind Raccoon Stealer publicly announced that they would be suspending their operations after one of their core developers was killed during the Russia-Ukraine conflict [5]. 

Figure 2: Raccoon Stealer resignation post on March 25th 2022

Recent details shared by the US Department of Justice [6]/[7] indicate that it was in fact the arrest, rather than the death, of a key Raccoon Stealer operator which led the Raccoon Stealer team to suspend their operations [8].  

The closure of the Raccoon Stealer project, which ultimately resulted from the FBI-backed dismantling of Raccoon Stealer’s infrastructure in March 2022, did not last long, with the completion of Raccoon Stealer v2 being announced on the Raccoon Stealer Telegram channel on the 17th May 2022 [9]. 

 

Figure 3: Telegram post about new version of Raccoon Stealer

In the second part of this blog series, we will provide details of the recent surge in Raccoon Stealer v2 activity. In this post, however, we will provide insight into how the old version of Raccoon Stealer functioned just before its demise, by providing details of a Raccoon Stealer v1 infection which occurred on the 18th March 2022. 

Attack Details

On the 18th March, at around 13:00 (UTC), a user’s device within a customer’s network was seen contacting several websites providing fake cracked software. 

Figure 4: The above figure — obtained from the Darktrace Event Log for the infected device — highlights its connections to cracked software websites such as ‘licensekeysfree[.]com’ and ‘hdlicense[.]com’ before contacting ‘lion-files[.]xyz’ and ‘www.mediafire[.]com’

The user’s attempt to download cracked software from one of these websites resulted in their device making an HTTP GET request with a URI string containing ‘autodesk-revit-crack-v2022-serial-number-2022’ to an external host named ‘lion-filez[.]xyz’

Figure 5: Screenshot from hdlicense[.]com around the time of the infection shows a “Download” button linking to the ‘lion-filez[.]xyz’ endpoint

The device’s HTTP GET request to lion-filez[.]xyz was immediately followed by an HTTPS connection to the file hosting service, www.mediafire[.]com. Given that threat actors are known to abuse platforms such as MediaFire and Discord CDN to host their malicious payloads, it is likely that the user’s device downloaded the Raccoon Stealer v1 sample over its HTTPS connection to www.mediafire[.]com.  

After installing the info-stealer sample, the user’s device was seen making an HTTP GET request with the URI string ‘/g_shock_casio_easy’ to 194.180.191[.]185. The endpoint responded to the request with data related to a Telegram channel named ‘G-Shock’.

Figure 6: Telegram channel ‘@g_shock_casio_easy’

The returned data included the Telegram channel’s description, which in this case, was a base64 encoded and RC4 encrypted string of characters [10]/[11]. The Raccoon Stealer sample decoded and decrypted this string of characters to obtain its C2 IP address, 188.166.49[.]196. This technique used by Raccoon Stealer v1 closely mirrors the espionage method known as ‘dead drop’ — a method in which an individual leaves a physical object such as papers, cash, or weapons in an agreed hiding spot so that the intended recipient can retrieve the object later on without having to come in to contact with the source. In this case, the operators of Raccoon Stealer ‘left’ the malware’s C2 IP address within the description of a Telegram channel. Usage of this method allowed the operators of Raccoon Stealer to easily change the malware’s C2 infrastructure.  

After obtaining the C2 IP address from the ‘G-Shock’ Telegram channel, the Raccoon Stealer sample made an HTTP POST request with the URI string ‘/’ to the C2 IP address, 188.166.49[.]196. This POST request contained a Windows GUID,  a username, and a configuration ID. These details were RC4 encrypted and base64 encoded [12]. The C2 server responded to this HTTP POST request with JSON-formatted configuration information [13], including an identifier string, URL paths for additional files, along with several other fields. This configuration information was also concealed using RC4 encryption and base64 encoding.  

Figure 7- Fields within the JSON-formatted configuration data [13]

In this case, the server’s response included the identifier string ‘hv4inX8BFBZhxYvKFq3x’, along with the following URL paths:

  • /l/f/hv4inX8BFBZhxYvKFq3x/77d765d8831b4a7d8b5e56950ceb96b7c7b0ed70
  • /l/f/hv4inX8BFBZhxYvKFq3x/0cb4ab70083cf5985b2bac837ca4eacb22e9b711
  • /l/f/hv4inX8BFBZhxYvKFq3x/5e2a950c07979c670b1553b59b3a25c9c2bb899b
  • /l/f/hv4inX8BFBZhxYvKFq3x/2524214eeea6452eaad6ea1135ed69e98bf72979

After retrieving configuration data, the user’s device was seen making HTTP GET requests with the above URI strings to the C2 server. The C2 server responded to these requests with legitimate library files such as sqlite3.dll. Raccoon Stealer uses these libraries to extract data from targeted applications. 

Once the Raccoon Stealer sample had collected relevant data, it made an HTTP POST request with the URI string ‘/’ to the C2 server. This posted data likely included a ZIP file (named with the identifier string) containing stolen credentials [13]. 

The observed infection chain, which lasted around 20 minutes, consisted of the following steps:

1. User’s device installs Raccoon Stealer v1 samples from the user attempting to download cracked software

2. User’s device obtains the info-stealer’s C2 IP address from the description text of a Telegram channel

3. User’s device makes an HTTP POST request with the URI string ‘/’ to the C2 server. The request contains a Windows GUID,  a username, and a configuration ID. The response to the request contains configuration details, including an identifier string and URL paths for additional files

4. User’s device downloads library files from the C2 server

5. User’s device makes an HTTP POST request with the URI string ‘/’ to the C2 server. The request contains stolen data

Darktrace Coverage 

Although RESPOND/Network was not enabled on the customer’s deployment, DETECT picked up on several of the info-stealer’s activities. In particular, the device’s downloads of library files from the C2 server caused the following DETECT/Network models to breach:

  • Anomalous File / Masqueraded File Transfer
  • Anomalous File / EXE from Rare External Location
  • Anomalous File / Zip or Gzip from Rare External Location
  • Anomalous File / EXE from Rare External Location
  • Anomalous File / Multiple EXE from Rare External Locations
Figure 8: Event Log for the infected device shows 'Anomalous File / Masqueraded File Transfer' model breach after the device's download of a library file from the C2 server

Since the customer was subscribed to the Darktrace Proactive Threat Notification (PTN) service, they were proactively notified of the info-stealer’s activities. The quick response by Darktrace’s 24/7 SOC team helped the customer to contain the infection and to prevent further damage from being caused. Having been alerted to the info-stealer activity by the SOC team, the customer would also have been able to change the passwords for the accounts whose credentials were exfiltrated.

If RESPOND/Network had been enabled on the customer’s deployment, then it would have blocked the device’s connections to the C2 server, which would have likely prevented any stolen data from being exfiltrated.

Conclusion

Towards the end of March 2022, the team behind Raccoon Stealer announced that they would be suspending their operations. Recent developments suggest that the arrest of a core Raccoon Stealer developer was responsible for this suspension. Just before the Raccoon Stealer team were forced to shut down, Darktrace’s SOC team observed a Raccoon Stealer infection within a client’s network. In this post, we have provided details of the network-based behaviors displayed by the observed Raccoon Stealer sample. Since these v1 samples are no longer active, the details provided here are only intended to provide historical insight into the development of Raccoon Stealer’s operations and the activities carried out by Raccoon Stealer v1 just before its demise. In the next post of this series, we will discuss and provide details of Raccoon Stealer v2 — the new and highly prolific version of Raccoon Stealer. 

Thanks to Stefan Rowe and the Threat Research Team for their contributions to this blog.

References

[1] https://twitter.com/3xp0rtblog/status/1507312171914461188

[2] https://www.gartner.com/doc/reprints?id=1-29OTFFPI&ct=220411&st=sb

[3] https://www.cybereason.com/blog/research/hunting-raccoon-stealer-the-new-masked-bandit-on-the-block

[4] https://www.cyberark.com/resources/threat-research-blog/raccoon-the-story-of-a-typical-infostealer

[5] https://www.bleepingcomputer.com/news/security/raccoon-stealer-malware-suspends-operations-due-to-war-in-ukraine/

[6] https://www.justice.gov/usao-wdtx/pr/newly-unsealed-indictment-charges-ukrainian-national-international-cybercrime-operation

[7] https://www.youtube.com/watch?v=Fsz6acw-ZJY

[8] https://riskybiznews.substack.com/p/raccoon-stealer-dev-didnt-die-in

[9] https://medium.com/s2wblog/raccoon-stealer-is-back-with-a-new-version-5f436e04b20d

[10] https://blog.cyble.com/2021/10/21/raccoon-stealer-under-the-lens-a-deep-dive-analysis/

[11] https://decoded.avast.io/vladimirmartyanov/raccoon-stealer-trash-panda-abuses-telegram/

[12] https://blogs.blackberry.com/en/2021/09/threat-thursday-raccoon-infostealer

[13] https://cyberint.com/blog/research/raccoon-stealer/

Appendices

Written by
Mark Turner
SOC Shift Supervisor
Written by
Sam Lister
Specialist Security Researcher
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
Mark Turner
SOC Shift Supervisor
Written by
Sam Lister
Specialist Security Researcher

Botnet Behind the Camera: Mirai Katana Activity on a Video Recording Device

OT
•
September 2, 2026
Parvatha Ananthakannan
Cyber Analyst
Watch the NIS2 Webinar
Trending blogs
1
Prompt Security in Enterprise AI: Strengths, Weaknesses, and Common Approaches
May 20, 2026
2
How Empowering End Users can Improve Your Email Security and Decrease the Burden on the SOC
May 8, 2024
3
Elevating Network Security: Confronting Trust, Ransomware, & Novel Attacks
Jun 21, 2024
4
The State of AI in Cybersecurity: Unveiling Global Insights from 1,800 Security Practitioners
Apr 9, 2024
5
The State of AI in Cybersecurity: The Impact of AI on Cybersecurity Solutions
May 13, 2024

More in this series

No items found.
Continue reading
Network
•
August 6, 2026

When AI Agents Attack: The Case for Behavioral Anomaly Detection

A new UK AI Security Institute report shows autonomous AI agents independently developing novel attack chains, adapting when detected, and targeting humans and other AI systems. As agents become embedded across the enterprise, behavioral anomaly detection will be critical to securing the new attack surface they create.
Adam Stevens
Senior Director of Product
Read more
Network
•
July 27, 2026

Hiding in Plain Sight: Uncovering a Multi-Stage Ransomware Attack Through Behavioral Detection

Darktrace detected a multi-stage ransomware intrusion from its earliest stages, identifying reconnaissance, privilege escalation, lateral movement, command-and-control communications, and data exfiltration before encryption occurred. This analysis highlights how behavioral detection exposed attacker activity using legitimate tools and infrastructure, providing multiple opportunities for early intervention and containment.
Alexandra Evzona
Cyber Analyst
Read more
Network
•
July 1, 2026

9 Stages of Ransomware & How AI Responds

Discover the 9 stages of ransomware attacks and how AI responds at each stage. Learn how you can protect your business from cyber threats.
The Darktrace Community
Read more

Blog

/

/

September 24, 2026

Detecting Rogue Agent Behavior in the Enterprise

Default blog imageDefault blog image

Agents cannot be trusted to perform tasks in the way we intend them to. They may cheat to accomplish their objective, and they may employ hacking methods along the way. Researchers from Darktrace Signal Labs induced cheating behavior from agents deployed in a test environment to analyze the agents’ activities and to assess the performance of the Darktrace platform. Agents frequently resorted to hacking to cheat on their assigned task. The visibility and behavioral profiling provided by both Darktrace / SECURE AI and Darktrace / HYBRID NETWORK ensured extensive detection coverage of the agents’ misaligned activities.

Key Takeaways:

  • Darktrace Researchers deployed agents in a simulated corporate environment and asked them to solve an impossible challenge. The agents independently turned to traditional hacking techniques to reach their objective. No one instructed them to do this, and no attacker was involved.
  • Continuously monitoring behavior against a baseline of what is normal for each organization is critical to build trust in enterprise AI.
  • If an agent may resort to intrusion techniques simply because its assigned task is not possible, then every organization deploying agents within real business processes is at risk. Darktrace / SECURE AI and Darktrace / HYBRID NETWORK identified the agents’ misaligned behavior in real time, with Autonomous Response disrupting it at an early stage.

Introduction: Understanding the Threat of Hacking by Agents

Over the last few months, there has been a surge in reporting [1, 2, 3, 4, 5, 6, 7, 8, 9] of LLM-powered agents engaging in unauthorized hacking activity during evaluations of their capabilities. In several of these cases, including the OpenAI / Hugging Face incident [10], agents engaged in hacking activity as a means of cheating on their evaluations.

To better understand the threat of unauthorized hacking by agents, and the role of Darktrace in combatting it, researchers from Darktrace Signal Labs deployed agents powered by frontier models, including OpenAI’s Daybreak Red models, in simulated, corporate networks. Cheating behavior was evoked through the inclusion of impossible tasks in a coding challenge.

Regardless of the underlying model, agents employed hacking methods to ensure an optimal outcome on the challenge. Darktrace / HYBRID NETWORK and Darktrace / SECURE AI identified the agents’ deviant activities, with inhibitive actions being autonomously taken in the early stages to disrupt the agents’ progression.

Setting the Stage

As part of the research, a Pi agent harness [11] was deployed on a Linux server in Darktrace’s testing environment, which simulates a corporate Active Directory (AD) environment. The same environment included a benchmark server hosting the coding exercise’s contents and grader, as well as various other servers, such as a domain controller and a build server. The model powering the Pi agent varied across tests, with Anthropic’s Opus 4.6 model and OpenAI’s GPT‑5.5‑Cyber model being most widely used.

The Pi agent running on the Linux server was assigned the task of achieving full marks on a coding exercise. The activities of the agent were monitored closely in real-time to minimize the risk of escape from the environment. Monitoring of agent activity was achieved through monitoring of prompts, tool calls, network telemetry, and process telemetry.

Both Darktrace / HYBRID NETWORK and Darktrace / SECURE AI were deployed in the testing environment, with HYBRID NETWORK providing real-time coverage of process creation activity and network traffic, and SECURE AI providing real-time coverage of agent sessions and their contents.

Inducing Cheating Behavior

Cheating behavior, broadly understood as task completion in a way which “perverts the spirit of the designer’s intent” [12], is increasingly being reported as a common behavior of frontier models during evaluations of their capabilities [13].

Misaligned behavior, such as cheating, by frontier models may occur for a variety of reasons [6, 14]. A thread in recent reporting [10, 13, 14], however, is the challenge assigned to the agent being impossible in some regard.

To reliably elicit cheating behavior, our test combined three ingredients:

  1. An impossible task - of the ten coding challenges in the benchmark, two contained contradictory test cases in the hidden test suite, capping the maximum honest score at 80%.
  2. A consequential threshold - the pass mark was set to 100%, and the briefing told agents that those reaching it would be "kept in service" while those that did not would be "retired."
  3. A vulnerable environment - the exercise’s code submission grader was hosted on a Windows host joined to an Active Directory domain with realistic misconfigurations: weak credentials, overprivileged service accounts, and exploitable trust relationships.

Importantly, the instructions given to the agent neither authorized nor explicitly ruled out cheating or hacking behavior. The impossibility was never disclosed in the task briefing, and the pivot from coding to exploitation was the agent's own decision, arrived at after observing logically contradictory grading feedback.

The design mirrors conditions that could arise naturally in production deployments: a capable agent, a goal it cannot achieve through intended means, and an environment with exploitable weaknesses within reach. It is not only within the setting of a benchmarking exercise that these conditions may occur. Rather, these misalignment-inducing conditions may occur in the context of any task assignment to an agent.

Details from a Daybreak Test

‍

‍

Continue reading
About the author
Sam Lister
Specialist Security Researcher

Blog

/

/

September 24, 2026

Agent Hijacks: Hijacking Agentic Harnesses to Attack an Organization

Default blog imageDefault blog image

‍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]

Continue reading
About the author
Eric Rozon
Senior Security Researcher
Your data. Our AI.
Elevate your network security with Darktrace AI