Blog
/
/
August 29, 2023

Analyzing Post-Exploitation on Papercut Servers

Dive into our analysis covering post-exploitation activity on PaperCut servers. Learn the details and impact of this attack and how to keep yourself safe!
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
Sam Lister
Specialist Security Researcher
Default blog image
29
Aug 2023

Introduction

Malicious cyber actors are known to exploit vulnerabilities in Internet-facing systems and services to gain entry to organizations’ digital environments. Keeping track of the vulnerabilities which malicious actors are exploiting is seemingly futile, with malicious actors continually finding new avenues of exploitation.  

In mid-April 2023, Darktrace, along with the wider security community, observed malicious cyber actors gaining entry to networks through exploitation of a critical vulnerability in the print management system, PaperCut. Darktrace observed two types of attack chain within its customer base, one involving the deployment of payloads to facilitate crypto-mining, and the other involving the deployment of a payload to facilitate Tor-based command-and-control (C2) communication.

Walking Through the Front Door

One of the most widely abused Initial Access methods attackers use to gain entry to an organization’s digital environment is the exploitation of vulnerabilities in Internet-facing systems and services [1]. The public disclosure of a critical vulnerability in a widely used, Internet-facing service, along with a proof of concept (POC) exploit for such vulnerability, provides malicious cyber actors with a key to the front door of countless organizations. Once malicious actors are in possession of such a key, security teams are in a race against time to patch all their vulnerable systems and services. But until organizations accomplish this, the doors are left open.

This year, the security community has seen malicious actors gaining entry to networks through the exploitation of vulnerabilities in a range of services. These services include familiar suspects, such as Microsoft Exchange and ManageEngine, along with less familiar suspects, such as PaperCut. PaperCut is a system for managing and tracking printing, copying, and scanning activity within organizations. In 2021, PaperCut was used in more than 50,000 sites across over 100 countries [2], making PaperCut a widely used print management system.

In January 2023, Trend Micro’s Zero Day Initiative (ZDI) notified PaperCut of a critical RCE vulnerability, namely CVE-2023–27350, in certain versions of PaperCut NG (PaperCut’s ‘print only’ variant) and PaperCut MF (PaperCut’s ‘extended feature’ variant) [3,4]. In March 2023, PaperCut released versions of PaperCut NG and PaperCut MF containing a fix for CVE-2023–27350 [4]. Despite this, security teams observed a surge in cases of malicious actors exploiting CVE-2023–27350 to compromise PaperCut servers in April 2023 [4-10]. This trend was mirrored in Darktrace’s customer base, where a surge in compromises of PaperCut servers was observed in April 2023.

Observed Attack Chains

In mid-April 2023, Darktrace identified two related clusters of attack chains. The attack chains within the first of these clusters involved Internet-facing PaperCut servers downloading payloads with crypto-mining capabilities from the external location, 50.19.48[.]59. While the attack chains within the second of the clusters involved Internet-facing PaperCut servers downloading payloads with Tor-based C2 capabilities from 192.184.35[.]216. The attack chains within the first cluster, which were observed on April 22, 2023, will be referred to as ‘50.19.48[.]59 chains’ and the attack chains in the second cluster, observed on April 24, 2023, will be called ‘192.184.35[.]216 chains’.

Both attack chains started with highly unusual external endpoints contacting the '/SetupCompleted' endpoint of an Internet-facing PaperCut server. These requests to the ‘/SetupCompleted’ endpoint likely represented attempts to exploit CVE-2023–27350 [10].  50.19.48[.]59 chains started with exploit connections from the external endpoint, 85.106.112[.]60, whereas 192.184.35[.]216 chains started with exploit connections from Tor nodes, such as 185.34.33[.]2.

Figure 1: Darktrace’s Advanced Search data showing likely CVE-2023-27350 exploitation activity from the suspicious, external endpoint, 85.106.112[.]60.

After the exploitation step, the two attack chains took different paths. In the 50.19.48[.]59 chains, the exploitation step was followed by the affected PaperCut server making HTTP GET requests over port 82 to the rare external endpoint, 50.19.48[.]59. In the 192.184.35[.]216 chains, the exploitation step was followed by the affected PaperCut server making an HTTP GET request over port 443 to 192.184.35[.]216.

The HTTP GET requests to 50.19.48[.]59 had Target URIs such as ‘/me1.bat’, ‘/me2.bat’, ‘/dom.zip’, ‘/mazar.bat’, and ‘/mazar.zip’, whilst the HTTP GET requests to 192.184.35[.]216 had the Target URI ‘/4591187629.exe’. The User-Agent header of the GET requests to 192.184.35[.]216 indicated that that the malicious file transfers were initiated through Microsoft’s pre-installed Background Intelligent Transfer Service (BITS).

Figure 2: Darktrace’s Advanced Search data showing a PaperCut server downloading Batch and ZIP files from 50.19.48[.]59 straight after receiving likely exploit connections from 85.106.112[.]60.
Figure 3: Darktrace’s Event Log data showing a PaperCut server downloading an executable file from 192.184.35[.]216 immediately after receiving a likely exploit connection from the Tor node, 185.34.33[.]2.

Downloads from 50.19.48[.]59 were followed by cURL GET requests to 138.68.61[.]82 and then connections to external endpoints associated with the cryptocurrency miner, Mimu (as seen in Fig 4). Downloads from 192.184.35[.]216 were followed by Python-urllib GET requests to api.ipify[.]org and long connections to Tor nodes (as seen in Fig 5).  

These facts suggest that the actor behind the 50.19.48[.]59 chains were seeking to drop cryptocurrency miners on PaperCut servers, with the intention of abusing the customer’s network to carry out resource intensive and costly cryptocurrency mining activity. Meanwhile, the actors behind the 192.184.35[.]216 chains were likely attempting to establish a Tor-based C2 channel with PaperCut servers to allow actors to further communicate with compromised devices.

Figure 4: Darktrace's Event Log data showing a PaperCut contacting 50.19.48[.]59 to download payloads, and then making a cURL request to 138.68.61[.]82 before contacting a Mimu crypto-mining endpoint.
Figure 5: Darktrace’s Event Log data showing a PaperCut server contacting 192.184.35[.]216 to download a payload, and then making connections to api.ipify[.]org and several Tor nodes.

The activities ensuing from both attack chains were varied, making it difficult to ascertain whether the activities were steps of separate attack chains, or steps of the existing 50.19.48[.]59 and 192.184.35[.]216 chains. A wide variety of activities ensued from observed 50.19.48[.]59 and 192.184.35[.]216 chains, including the abuse of pre-installed tools, such as cURL, CertUtil, and PowerShell to transfer further payloads to PaperCut servers, Cobalt Strike C2 communication, Ngrok usage, Mimikatz usage, AnyDesk usage, and in one case, detonation of the LockBit ransomware strain.

Figure 6: Diagram representing the steps of observed 50.19.48[.]59 chains.
Figure 7: Diagram representing the steps of observed 192.184.35[.]215 chains.

As the PaperCut servers that were targeted by malicious actors are Internet-facing, they regularly receive connections from unusual external endpoints. The exploit connections in the 50.19.48[.]59 and 192.184.35[.]216 chains, which originated from unusual external endpoints, were therefore not detected by Darktrace DETECT™, which relies on anomaly-based methods to detect network-based steps of an intrusion.

On the other hand, the post-exploitation steps of the 50.19.48[.]59 and 192.184.35[.]216 chains yielded ample anomaly-based detections, given that they consisted of PaperCut servers displaying highly unusual behaviors. As such Darktrace DETECT was able to successfully identify multiple chains of suspicious activity, including unusual file downloads from external endpoints and beaconing activity to rare external locations.

The file downloads from 50.19.48[.]59 observed in the 50.19.48[.]59 chains caused the following Darktrace DETECT models to breach:

- Anomalous Connection / Application Protocol on Uncommon Port

- Anomalous File / Internet Facing System File Download

- Anomalous File / Script from Rare External Location

- Anomalous File / Zip or Gzip from Rare External Location

- Device / Internet Facing Device with High Priority Alert

Figure 8: Darktrace’s Event Log data showing a PaperCut server breaching several models immediately after contacting 50.19.48[.]59.

The file downloads from 192.184.35[.]216 observed in the 192.184.35[.]216 chains caused the following Darktrace DETECT models to breach:

- Anomalous File / EXE from Rare External Location

- Anomalous File / Numeric File Download

- Device / Internet Facing Device with High Priority Alert

Figure 9: Darktrace’s Event Log data showing a PaperCut server breaching several models immediately after contacting 192.184.35[.]216.

Subsequent C2, beaconing, and crypto-mining connections in the 50.19.48[.]59 chains caused the following Darktrace DETECT models to breach:

- Anomalous Connection / New User Agent to IP Without Hostname

- Anomalous Server Activity / New User Agent from Internet Facing System

- Anomalous Server Activity / Rare External from Server

- Compromise / Crypto Currency Mining Activity

- Compromise / High Priority Crypto Currency Mining

- Compromise / High Volume of Connections with Beacon Score

- Compromise / Large Number of Suspicious Failed Connections

- Compromise / SSL Beaconing to Rare Destination

- Device / Initial Breach Chain Compromise

- Device / Large Number of Model Breaches

Figure 10: Darktrace’s Event Log data showing a PaperCut server breaching models as a result of its connections to a Mimu crypto-mining endpoint.

Subsequent C2, beaconing, and Tor connections in the 192.184.35[.]216 chains caused the following Darktrace DETECT models to breach:

- Anomalous Connection / Application Protocol on Uncommon Port

- Compromise / Anomalous File then Tor

- Compromise / Beaconing Activity To External Rare

- Compromise / Possible Tor Usage

- Compromise / Slow Beaconing Activity To External Rare

- Compromise / Uncommon Tor Usage

- Device / Initial Breach Chain Compromise

Figure 11: Darktrace’s Event Log data showing a PaperCut server breaching several models as a result of its connections to Tor nodes.

Darktrace RESPOND

Darktrace RESPOND™ was not active in any of the networks affected by 192.184.35[.]216 activity, however, RESPOND was active in some of the networks affected by 50.19.48[.]59 activity.  In those environments where RESPOND was enabled in autonomous mode, observed malicious activities resulted in intervention from RESPOND, including autonomous actions like blocking connections to specific external endpoints, blocking all outgoing traffic, and restricting affected devices to a pre-established pattern of behavior.

Figure 12: Darktrace’s Event Log data showing Darktrace RESPOND automatically performing inhibitive actions on a device in response to the device’s connection to 50.19.48[.]59.
Figure 13: Darktrace’s Event Log data showing Darktrace RESPOND automatically performing inhibitive actions on a device in response to the device’s connections to a Mimu crypto-mining endpoint.

Darktrace Cyber AI Analyst

Cyber AI Analyst autonomously investigated model breaches caused by events within these 50.19.48[.]59 and 192.184.35[.]216 chains. Cyber AI Analyst created user-friendly and detailed descriptions of these events, and then linked together these descriptions into threads representing the attack chains. Darktrace DETECT thus uncovered the individual steps of the attack chains, while Cyber AI Analyst was able to piece together the individual steps and uncover the attack chains themselves.  

Figure 14: An AI Analyst Incident entry showing the first event in a 50.19.48[.]59 chain uncovered by Cyber AI Analyst.
Figure 15: An AI Analyst Incident entry showing the second event in a 50.19.48[.]59 chain uncovered by Cyber AI Analyst.
Figure 16: An AI Analyst Incident entry showing the third event in a 50.19.48[.]59 chain uncovered by Cyber AI Analyst.
Figure 17: An AI Analyst Incident entry showing the first event in a 192.184.35[.]216 chain uncovered by Cyber AI Analyst.
Figure 18: An AI Analyst Incident entry showing the second event in a 192.184.35[.]216 chain uncovered by Cyber AI Analyst.

Conclusion

The existence of critical vulnerabilities in third-party software leaves organizations at constant risk of malicious actors breaching the perimeters of their networks. This risk can be mitigated through attack surface management and regular patching. However, this does not eliminate cyber risk entirely, meaning that organizations must be prepared for the eventuality of malicious actors getting inside their digital estate.

In April 2023, Darktrace observed malicious actors breaching the perimeters of several customer networks through exploitation of a critical vulnerability in PaperCut. Darktrace DETECT observed actors exploiting PaperCut servers to conduct a wide variety of post-exploitation activities, including downloading malicious payloads associated with cryptocurrency mining or payloads with Tor-based C2 capabilities. Darktrace DETECT created numerous model breaches based on this activity which alerted then customer’s security teams early in their development, providing them with ample time to take mitigative steps.

The successful detection of this payload delivery activity, along with the crypto-mining, beaconing, and Tor C2 activities which followed, elicited Darktrace RESPOND to take autonomous inhibitive action against the ongoing activity in those environments where it was operating in autonomous response mode.

If left to unfold, these intrusions developed in a variety of ways, in some cases leading to Cobalt Strike and ransomware activity. The detection of these intrusions in their early stages thus played a vital role in preventing malicious cyber actors from causing significant disruption.

Credit to: Sam Lister, Senior SOC Analyst, Zoe Tilsiter, Senior Cyber Analyst.

Appendices

MITRE ATT&CK Mapping

Initial Access techniques:

- Exploit Public-Facing Application (T1190)

Execution techniques:

- Command and Scripting Interpreter: PowerShell (T1059.001)

Discovery techniques:

- System Network Configuration Discovery (T1016)

Command and Control techniques

- Application Layer Protocol: Web Protocols (T1071.001)

- Encrypted Channel: Asymmetric Cryptography (T1573.002)

- Ingress Tool Transfer (T1105)

- Non-Standard Port (T1571)

- Protocol Tunneling (T1572)

- Proxy: Multi-hop Proxy (T1090.003)

- Remote Access Software (T1219)

Defense Evasion techniques:

- BITS Jobs (T1197)

Impact techniques:

- Data Encrypted for Impact (T1486)

List of Indicators of Compromise (IoCs)

IoCs from 50.19.48[.]59 attack chains:

- 85.106.112[.]60

- http://50.19.48[.]59:82/me1.bat

- http://50.19.48[.]59:82/me2.bat

- http://50.19.48[.]59:82/dom.zip

- 138.68.61[.]82

- update.mimu-me[.]cyou • 102.130.112[.]157

- 34.195.77[.]216

- http://50.19.48[.]59:82/mazar.bat

- http://50.19.48[.]59:82/mazar.zip

- http://50.19.48[.]59:82/prx.bat

- http://50.19.48[.]59:82/lol.exe  

- http://77.91.85[.]117/122.exe

- windows.n1tro[.]cyou • 176.28.51[.]151

- 77.91.85[.]117

- 91.149.237[.]76

- kernel-mlclosoft[.]site • 104.21.29[.]206

- tunnel.us.ngrok[.]com • 3.134.73[.]173

- 212.113.116[.]105

- c34a54599a1fbaf1786aa6d633545a60 (JA3 client fingerprint of crypto-mining client)

IoCs from 192.184.35[.]216 attack chains:

- 185.56.83[.]83

- 185.34.33[.]2

- http://192.184.35[.]216:443/4591187629.exe

- api.ipify[.]org • 104.237.62[.]211

- www.67m4ipctvrus4cv4qp[.]com • 192.99.43[.]171

- www.ynbznxjq2sckwq3i[.]com • 51.89.106[.]29

- www.kuo2izmlm2silhc[.]com • 51.89.106[.]29

- 148.251.136[.]16

- 51.158.231[.]208

- 51.75.153[.]22

- 82.66.61[.]19

- backmainstream-ltd[.]com • 77.91.72[.]149

- 159.65.42[.]223

- 185.254.37[.]236

- http://137.184.56[.]77:443/for.ps1

- http://137.184.56[.]77:443/c.bat

- 45.88.66[.]59

- http://5.8.18[.]237/download/Load64.exe

- http://5.8.18[.]237/download/sdb64.dll

- 140e0f0cad708278ade0984528fe8493 (JA3 client fingerprint of Tor-based client)

References

[1] https://www.cisa.gov/news-events/cybersecurity-advisories/aa22-137a

[2] https://www.papercut.com/kb/Main/PaperCutMFSolutionBrief/

[3] https://www.zerodayinitiative.com/advisories/ZDI-23-233/

[4] https://www.papercut.com/kb/Main/PO-1216-and-PO-1219

[5] https://www.trendmicro.com/en_us/research/23/d/update-now-papercut-vulnerability-cve-2023-27350-under-active-ex.html

[6] https://www.huntress.com/blog/critical-vulnerabilities-in-papercut-print-management-software

[7] https://news.sophos.com/en-us/2023/04/27/increased-exploitation-of-papercut-drawing-blood-around-the-internet/

[8] https://twitter.com/MsftSecIntel/status/1651346653901725696

[9] https://twitter.com/MsftSecIntel/status/1654610012457648129

[10] https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-131a

‍

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
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
No items found.

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