Blog
/
Network
/
April 2, 2024

Darktrace's Investigation of Raspberry Robin Worm

Discover how Darktrace is leading the hunt for Raspberry Robin. Explore early insights and strategies in the battle against cyber threats.
Written by
Alexandra Sentenac
Cyber Analyst
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
Alexandra Sentenac
Cyber Analyst
Default blog image
02
Apr 2024

Introduction

In the face of increasingly hardened digital infrastructures and skilled security teams, malicious actors are forced to constantly adapt their attack methods, resulting in sophisticated attacks that are designed to evade human detection and bypass traditional network security measures.  

One such example that was recently investigated by Darktrace is Raspberry Robin, a highly evasive worm malware renowned for merging existing and novel techniques, as well as leveraging both physical hardware and software, to establish a foothold within organization’s networks and propagate additional malicious payloads.

What is Raspberry Robin?

Raspberry Robin, also known as ‘QNAP worm’, is a worm malware that was initially discovered at the end of 2023 [1], however, its debut in the threat landscape may have predated this, with Microsoft uncovering malicious artifacts linked to this threat (which it tracks under the name Storm-0856) dating back to 2019 [4]. At the time, little was known regarding Raspberry Robin’s objectives or operators, despite the large number of successful infections worldwide. While the identity of the actors behind Raspberry Robin still remains a mystery, more intelligence has been gathered about the malware and its end goals as it was observed delivering payloads from different malware families.

Who does Raspberry Robin target?

While it was initially reported that Raspberry Robin primarily targeted the technology and manufacturing industries, researchers discovered that the malware had actually targeted multiple sectors [3] [4]. Darktrace’s own investigations echoed this, with Raspberry Robin infections observed across various industries, including public administration, finance, manufacturing, retail education and transportation.

How does Raspberry Robin work?

Initially, it appeared that Raspberry Robin's access to compromised networks had not been utilized to deliver final-stage malware payloads, nor to steal corporate data. This uncertainty led researchers to question whether the actors involved were merely “cybercriminals playing around” or more serious threats [3]. This lack of additional exploitation was indeed peculiar, considering that attackers could easily escalate their attacks, given Raspberry Robin’s ability to bypass User Account Control using legitimate Windows tools [4].

However, at the end of July 2022, some clarity emerged regarding the operators' end goals. Microsoft researchers revealed that the access provided by Raspberry Robin was being utilized by an access broker tracked as DEV-0206 to distribute the FakeUpdates malware downloader [2]. Researchers further discovered malicious activity associated with Evil Corp TTPs (i.e., DEV-0243) [5] and payloads from the Fauppod malware family leveraging Raspberry Robin’s access [8]. This indicates that Raspberry Robin may, in fact, be an initial access broker, utilizing its presence on hundreds of infected networks to distribute additional payloads for paying malware operators. Thus far, Raspberry Robin has been observed distributing payloads linked to FIN11, Clop Gang, BumbleBee, IcedID, and TrueBot on compromised networks [12].

Raspberry Robin’s Continued Evolution

Since it first appeared in the wild, Raspberry Robin has evolved from "being a widely distributed worm with no observed post-infection actions [...] to one of the largest malware distribution platforms currently active" [8]. The fact that Raspberry Robin has become such a prevalent threat is likely due to the continual addition of new features and evasion capabilities to their malware [6] [7].  

Since its emergence, the malware has “changed its communication method and lateral movement” [6] in order to evade signature detections based on threat intelligence and previous versions. Endpoint security vendors commonly describe it as heavily obfuscated malware, employing multiple layers of evasion techniques to hinder detection and analysis. These include for example dropping a fake payload when analyzed in a sandboxed environment and using mixed-case executing commands, likely to avoid case-sensitive string-based detections.  

In more recent campaigns, Raspberry Robin further appears to have added a new distribution method as it was observed being downloaded from archive files sent as attachments using the messaging service Discord [11]. These attachments contained a legitimate and signed Windows executable, often abused by attackers for side-loading, alongside a malicious dynamic-link library (DLL) containing a Raspberry Robin sample.

Another reason for the recent success of the malware may be found in its use of one-day exploits. According to researchers, Raspberry Robin now utilizes several local privilege escalation exploits that had been recently disclosed, even before a proof of concept had been made available [9] [10]. This led cyber security professionals to believe that operators of the malware may have access to an exploit seller [6]. The use of these exploits enhances Raspberry Robin's detection evasion and persistence capabilities, enabling it to propagate on networks undetected.

Darktrace’s Coverage of Raspberry Robin

Through two separate investigations carried out by Darktrace’s Threat Research team, first in late 2022 and then in November 2023, it became evident that Raspberry Robin was capable of integrating new functionalities and tactics, techniques and procedures (TTPs) into its attacks. Darktrace DETECT™ provided full visibility over the evolving campaign activity, allowing for a comparison of the threat across both investigations. Additionally, if Darktrace RESPOND™ was enabled on affected networks, it was able to quickly mitigate and contain emerging activity during the initial stages, thwarting the further escalation of attacks.

Raspberry Robin Initial Infection

The most prevalent initial infection vector appears to be the introduction of an infected external drive, such as a USB stick, containing a malicious .LNK file (i.e., a Windows shortcut file) disguised as a thumb drive or network share. When clicked, the LNK file automatically launches cmd.exe to execute the malicious file stored on the external drive, and msiexec.exe to connect to a Raspberry Robin command-and-control (C2) endpoint and download the main malware component. The whole process leverages legitimate Windows processes and is therefore less likely to raise any alarms from more traditional security solutions. However, Darktrace DETECT was able to identify the use of Msiexec to connect to a rare endpoint as anomalous in every case investigated.

Little is currently known regarding how the external drives are infected and distributed, but it has been reported that affected USB drives had previously been used for printing at printing and copying shops, suggesting that the infection may have originated from such stores [13].

A method as simple as leaving an infected USB on a desk in a public location can be a highly effective social engineering tactic for attackers. Exploiting both curiosity and goodwill, unsuspecting individuals may innocently plug in a found USB, hoping to identify its owner, unaware that they have unwittingly compromised their device.

As Darktrace primarily operates on the network layer, the insertion of a USB endpoint device would not be within its visibility. Nevertheless, Darktrace did observe several instances wherein multiple Microsoft endpoints were contacted by compromised devices prior to the first connection to a Raspberry Robin domain. For example, connections to the URI '/fwlink/?LinkID=252669&clcid=0x409' were observed in multiple customer environments prior to the first Raspberry Robin external connection. This connectivity seems to be related to Windows attempting to retrieve information about installed hardware, such as a printer, and could also be related to the inserting of an external USB drive.

Figure 1: Device Event Log showing an affected device making connections to Microsoft endpoints, prior to contacting the Raspberry Robin C2 endpoint ‘vqdn[.]net’.
Figure 1: Device Event Log showing an affected device making connections to Microsoft endpoints, prior to contacting the Raspberry Robin C2 endpoint ‘vqdn[.]net’.

Raspberry Robin Command-and-Control Activity

In all cases investigated by Darktrace, compromised devices were detected making HTTP GET connections via the unusual port 8080 to Raspberry Robin C2 endpoints using the new user agent 'Windows Installer'.

The C2 hostnames observed were typically short and matched the regex /[a-zA-Z0-9]{2,4}.[a-zA-Z0-9]{2,6}/, and were hosted on various top-level domains (TLD) such as ‘.rocks’, ‘.pm’, and ‘.wf’. On one customer network, Darktrace observed the download of an MSI file from the Raspberry Robin domain ‘wak[.]rocks’. This package contained a heavily protected malicious DLL file whose purpose was unknown at the time.  

However, in September 2022, external researchers revealed that the main purpose of this DLL was to download further payloads and enable lateral movement, persistence and privilege escalation on compromised devices, as well as exfiltrating sensitive information about the device. As worm infections spread through networks automatically, exfiltrating device data is an essential process for threat actor to keep track of which systems have been infected.

On affected networks investigated by Darktrace, compromised devices were observed making C2 connections that contained sensitive device information, including hostnames and credentials, with additional host information likely found within the data packets [12].

Figure 2: Model Breach Event Log displaying the events that triggered the the ‘New User Agent and Suspicious Request Data’ DETECT model breach.
Figure 2: Model Breach Event Log displaying the events that triggered the the ‘New User Agent and Suspicious Request Data’ DETECT model breach.

As for C2 infrastructure, Raspberry Robin leverages compromised Internet of Things (IoT) devices such as QNAP network attached storage (NAS) systems with hijacked DNS settings [13]. NAS devices are data storage servers that provide access to the files they store from anywhere in the world. These features have been abused by Raspberry Robin operators to distribute their malicious payloads, as any uploaded file could be stored and shared easily using NAS features.

However, Darktrace found that QNAP servers are not the only devices being exploited by Raspberry Robin, with DETECT identifying other IoT devices being used as C2 infrastructure, including a Cerio wireless access point in one example. Darktrace recognized that this connection was new to the environment and deemed it as suspicious, especially as it also used new software and an unusual port for the HTTP protocol (i.e., 8080 rather than 80).

In several instances, Darktrace observed Raspberry Robin utilizing TOR exit notes as backup C2 infrastructure, with compromised devices detected connecting to TOR endpoints.

Figure 3: Raspberry Robin C2 endpoint when viewed in a sandbox environment.
Figure 3: Raspberry Robin C2 endpoint when viewed in a sandbox environment.
Figure 4: Raspberry Robin C2 endpoint when viewed in a sandbox environment.
Figure 4: Raspberry Robin C2 endpoint when viewed in a sandbox environment.

Raspberry Robin in 2022 vs 2023

Despite the numerous updates and advancements made to Raspberry Robin between the investigations carried out in 2022 and 2023, Darktrace’s detection of the malware was largely the same.

DETECT models breached during first investigation at the end of 2022:

  • Device / New User Agent
  • Anomalous Server Activity / New User Agent from Internet Facing System
  • Device / New User Agent and New IP
  • Compromise / Suspicious Request Data
  • Compromise / Uncommon Tor Usage
  • Possible Tor Usage

DETECT models breached during second investigation in late 2023:

  • Device / New User Agent and New IP
  • Device / New User Agent and Suspicious Request Data
  • Device / New User Agent
  • Device / Suspicious Domain
  • Possible Tor Usage

Darktrace’s anomaly-based approach to threat detection enabled it to consistently detect the TTPs and IoCs associated with Raspberry Robin across the two investigations, despite the operator’s efforts to make it stealthier and more difficult to analyze.

In the first investigation in late 2022, Darktrace detected affected devices downloading addition executable (.exe) files following connections to the Raspberry Robin C2 endpoint, including a numeric executable file that appeared to be associated with the Vidar information stealer. Considering the advanced evasion techniques and privilege escalation capabilities of Raspberry Robin, early detection is key to prevent the malware from downloading additional malicious payloads.

In one affected customer environment investigated in late 2023, a total of 12 devices were compromised between mid-September and the end of October. As this particular customer did not have Darktrace RESPOND, the Raspberry Robin infection was able to spread through the network unabated until the customer acted upon Darktrace DETECT’s alerts.

Had Darktrace RESPOND been enabled in autonomous response mode, it would have been able to take immediate action following the first observed connection to a Raspberry Robin C2 endpoint, by blocking connections to the suspicious endpoint and enforcing a device’s normal ‘pattern of life’.

By enforcing a pattern of life on an affected device, RESPOND would prevent it from carrying out any activity that deviates from this learned pattern, including connections to new endpoints using new software as was the case in Figure 5, effectively shutting down the attack in the first instance.

Model Breach Event Log showing RESPOND’s actions against connections to Raspberry Robin C2 endpoints.
Figure 5: Model Breach Event Log showing RESPOND’s actions against connections to Raspberry Robin C2 endpoints.

Conclusion

Raspberry Robin is a highly evasive and adaptable worm known to evolve and change its TTPs on a regular basis in order to remain undetected on target networks for as long as possible. Due to its ability to drop additional malware variants onto compromised devices, it is crucial for organizations and their security teams to detect Raspberry Robin infections at the earliest possible stage to prevent the deployment of potentially disruptive secondary attacks.

Despite its continued evolution, Darktrace's detection of Raspberry Robin remained largely unchanged across the two investigations. Rather than relying on previous IoCs or leveraging existing threat intelligence, Darktrace DETECT’s anomaly-based approach allows it to identify emerging compromises by detecting the subtle deviations in a device’s learned behavior that would typically come with a malware compromise.

By detecting the attacks at an early stage, Darktrace gave its customers full visibility over malicious activity occurring on their networks, empowering them to identify affected devices and remove them from their environments. In cases where Darktrace RESPOND was active, it would have been able to take autonomous follow-up action to halt any C2 communication and prevent the download of any additional malicious payloads.  

Credit to Alexandra Sentenac, Cyber Analyst, Trent Kessler, Senior Cyber Analyst, Victoria Baldie, Director of Incident Management

Appendices

Darktrace DETECT Model Coverage

Device / New User Agent and New IP

Device / New User Agent and Suspicious Request Data

Device / New User Agent

Compromise / Possible Tor Usage

Compromise / Uncommon Tor Usage

MITRE ATT&CK Mapping

Tactic - Technique

Command & Control - T1090.003 Multi-hop Proxy

Lateral Movement - T1210 Exploitation of remote services

Exfiltration over C2 Data - T1041 Exfiltration over C2 Channel

Data Obfuscation - T1001 Data Obfuscation

Vulnerability Scanning - T1595.002 Vulnerability Scanning

Non-Standard Port - T1571 Non-Standard Port

Persistence - T1176 Browser Extensions

Initial Access - T1189 Drive By Compromise / T1566.002  Spearphishing Link

Collection - T1185 Man in the browser

List of IoCs

IoC - Type - Description + Confidence

vqdn[.]net - Hostname - C2 Server

mwgq[.]net - Hostname - C2 Server

wak[.]rocks - Hostname - C2 Server

o7car[.]com - Hostname - C2 Server

6t[.]nz - Hostname - C2 Server

fcgz[.]net - Hostname - Possible C2 Server

d0[.]wf - Hostname - C2 Server

e0[.]wf - Hostname - C2 Server

c4z[.]pl - Hostname - C2 Server

5g7[.]at - Hostname - C2 Server

5ap[.]nl - Hostname - C2 Server

4aw[.]ro - Hostname - C2 Server

0j[.]wf - Hostname - C2 Server

f0[.]tel - Hostname - C2 Server

h0[.]pm - Hostname - C2 Server

y0[.]pm - Hostname - C2 Server

5qy[.]ro - Hostname - C2 Server

g3[.]rs - Hostname - C2 Server

5qe8[.]com - Hostname - C2 Server

4j[.]pm - Hostname - C2 Server

m0[.]yt - Hostname - C2 Server

zk4[.]me - Hostname - C2 Server

59.15.11[.]49 - IP address - Likely C2 Server

82.124.243[.]57 - IP address - C2 Server

114.32.120[.]11 - IP address - Likely C2 Server

203.186.28[.]189 - IP address - Likely C2 Server

70.124.238[.]72 - IP address - C2 Server

73.6.9[.]83 - IP address - Likely C2 Server

References

[1] https://redcanary.com/blog/raspberry-robin/  

[2] https://www.bleepingcomputer.com/news/security/microsoft-links-raspberry-robin-malware-to-evil-corp-attacks/

[3] https://7095517.fs1.hubspotusercontent-na1.net/hubfs/7095517/FLINT%202022-016%20-%20QNAP%20worm_%20who%20benefits%20from%20crime%20(1).pdf

[4] https://www.bleepingcomputer.com/news/security/microsoft-finds-raspberry-robin-worm-in-hundreds-of-windows-networks/

[5] https://therecord.media/microsoft-ties-novel-raspberry-robin-malware-to-evil-corp-cybercrime-syndicate

[6] https://securityaffairs.com/158969/malware/raspberry-robin-1-day-exploits.html

[7] https://research.checkpoint.com/2024/raspberry-robin-keeps-riding-the-wave-of-endless-1-days/

[8] https://redmondmag.com/articles/2022/10/28/microsoft-details-threat-actors-leveraging-raspberry-robin-worm.aspx

[9] https://www.bleepingcomputer.com/news/security/raspberry-robin-malware-evolves-with-early-access-to-windows-exploits/

[10] https://www.bleepingcomputer.com/news/security/raspberry-robin-worm-drops-fake-malware-to-confuse-researchers/

[11] https://thehackernews.com/2024/02/raspberry-robin-malware-upgrades-with.html

[12] https://decoded.avast.io/janvojtesek/raspberry-robins-roshtyak-a-little-lesson-in-trickery/

[13] https://blog.bushidotoken.net/2023/05/raspberry-robin-global-usb-malware.html

Written by
Alexandra Sentenac
Cyber Analyst
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
Alexandra Sentenac
Cyber Analyst

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