ブログ
/
Network
/
April 24, 2024

How Cactus Ransomware Was Detected and Stopped

Discover the tactics used to uncover a Cactus ransomware infection and the implications for cybersecurity defenses.
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
Tiana Kelly
Senior Cyber Analyst & Team Lead
Default blog image
24
Apr 2024

What is CACTUS Ransomware?

In May 2023, Kroll Cyber Threat Intelligence Analysts identified CACTUS as a new ransomware strain that had been actively targeting large commercial organizations since March 2023 [1]. CACTUS ransomware gets its name from the filename of the ransom note, “cAcTuS.readme.txt”. Encrypted files are appended with the extension “.cts”, followed by a number which varies between attacks, e.g. “.cts1” and “.cts2”.

As the cyber threat landscape adapts to ever-present fast-paced technological change, ransomware affiliates are employing progressively sophisticated techniques to enter networks, evade detection and achieve their nefarious goals.

How does CACTUS Ransomware work?

In the case of CACTUS, threat actors have been seen gaining initial network access by exploiting Virtual Private Network (VPN) services. Once inside the network, they may conduct internal scanning using tools like SoftPerfect Network Scanner, and PowerShell commands to enumerate endpoints, identify user accounts, and ping remote endpoints. Persistence is maintained by the deployment of various remote access methods, including legitimate remote access tools like Splashtop, AnyDesk, and SuperOps RMM in order to evade detection, along with malicious tools like Cobalt Strike and Chisel. Such tools, as well as custom scripts like TotalExec, have been used to disable security software to distribute the ransomware binary. CACTUS ransomware is unique in that it adopts a double-extortion tactic, stealing data from target networks and then encrypting it on compromised systems [2].

At the end of November 2023, cybersecurity firm Arctic Wolf reported instances of CACTUS attacks exploiting vulnerabilities on the Windows version of the business analytics platform Qlik, specifically CVE-2023-41266, CVE-2023-41265, and CVE-2023-48365, to gain initial access to target networks [3]. The vulnerability tracked as CVE-2023-41266 can be exploited to generate anonymous sessions and perform HTTP requests to unauthorized endpoints, whilst CVE-2023-41265 does not require authentication and can be leveraged to elevate privileges and execute HTTP requests on the backend server that hosts the application [2].

Darktrace’s Coverage of CACTUS Ransomware

In November 2023, Darktrace observed malicious actors leveraging the aforementioned method of exploiting Qlik to gain access to the network of a customer in the US, more than a week before the vulnerability was reported by external researchers.

Here, Qlik vulnerabilities were successfully exploited, and a malicious executable (.exe) was detonated on the network, which was followed by network scanning and failed Kerberos login attempts. The attack culminated in the encryption of numerous files with extensions such as “.cts1”, and SMB writes of the ransom note “cAcTuS.readme.txt” to multiple internal devices, all of which was promptly identified by Darktrace DETECT™.

While traditional rules and signature-based detection tools may struggle to identify the malicious use of a legitimate business platform like Qlik, Darktrace’s Self-Learning AI was able to confidently identify anomalous use of the tool in a CACTUS ransomware attack by examining the rarity of the offending device’s surrounding activity and comparing it to the learned behavior of the device and its peers.

Unfortunately for the customer in this case, Darktrace RESPOND™ was not enabled in autonomous response mode during their encounter with CACTUS ransomware meaning that attackers were able to successfully escalate their attack to the point of ransomware detonation and file encryption. Had RESPOND been configured to autonomously act on any unusual activity, Darktrace could have prevented the attack from progressing, stopping the download of any harmful files, or the encryption of legitimate ones.

Cactus Ransomware Attack Overview

Holiday periods have increasingly become one of the favoured times for malicious actors to launch their attacks, as they can take advantage of the festive downtime of organizations and their security teams, and the typically more relaxed mindset of employees during this period [4].

Following this trend, in late November 2023, Darktrace began detecting anomalous connections on the network of a customer in the US, which presented multiple indicators of compromise (IoCs) and tactics, techniques and procedures (TTPs) associated with CACTUS ransomware. The threat actors in this case set their attack in motion by exploiting the Qlik vulnerabilities on one of the customer’s critical servers.

Darktrace observed the server device making beaconing connections to the endpoint “zohoservice[.]net” (IP address: 45.61.147.176) over the course of three days. This endpoint is known to host a malicious payload, namely a .zip file containing the command line connection tool PuttyLink [5].

Darktrace’s Cyber AI Analyst was able to autonomously identify over 1,000 beaconing connections taking place on the customer’s network and group them together, in this case joining the dots in an ongoing ransomware attack. AI Analyst recognized that these repeated connections to highly suspicious locations were indicative of malicious command-and-control (C2) activity.

Cyber AI Analyst Incident Log showing the offending device making over 1,000 connections to the suspicious hostname “zohoservice[.]net” over port 8383, within a specific period.
Figure 1: Cyber AI Analyst Incident Log showing the offending device making over 1,000 connections to the suspicious hostname “zohoservice[.]net” over port 8383, within a specific period.

The infected device was then observed downloading the file “putty.zip” over a HTTP connection using a PowerShell user agent. Despite being labelled as a .zip file, Darktrace’s detection capabilities were able to identify this as a masqueraded PuttyLink executable file. This activity resulted in multiple Darktrace DETECT models being triggered. These models are designed to look for suspicious file downloads from endpoints not usually visited by devices on the network, and files whose types are masqueraded, as well as the anomalous use of PowerShell. This behavior resembled previously observed activity with regards to the exploitation of Qlik Sense as an intrusion technique prior to the deployment of CACTUS ransomware [5].

The downloaded file’s URI highlighting that the file type (.exe) does not match the file's extension (.zip). Information about the observed PowerShell user agent is also featured.
Figure 2: The downloaded file’s URI highlighting that the file type (.exe) does not match the file's extension (.zip). Information about the observed PowerShell user agent is also featured.

Following the download of the masqueraded file, Darktrace observed the initial infected device engaging in unusual network scanning activity over the SMB, RDP and LDAP protocols. During this activity, the credential, “service_qlik” was observed, further indicating that Qlik was exploited by threat actors attempting to evade detection. Connections to other internal devices were made as part of this scanning activity as the attackers attempted to move laterally across the network.

Numerous failed connections from the affected server to multiple other internal devices over port 445, indicating SMB scanning activity.
Figure 3: Numerous failed connections from the affected server to multiple other internal devices over port 445, indicating SMB scanning activity.

The compromised server was then seen initiating multiple sessions over the RDP protocol to another device on the customer’s network, namely an internal DNS server. External researchers had previously observed this technique in CACTUS ransomware attacks where an RDP tunnel was established via Plink [5].

A few days later, on November 24, Darktrace identified over 20,000 failed Kerberos authentication attempts for the username “service_qlik” being made to the internal DNS server, clearly representing a brute-force login attack. There is currently a lack of open-source intelligence (OSINT) material definitively listing Kerberos login failures as part of a CACTUS ransomware attack that exploits the Qlik vulnerabilities. This highlights Darktrace’s ability to identify ongoing threats amongst unusual network activity without relying on existing threat intelligence, emphasizing its advantage over traditional security detection tools.

Kerberos login failures being carried out by the initial infected device. The destination device detected was an internal DNS server.
Figure 4: Kerberos login failures being carried out by the initial infected device. The destination device detected was an internal DNS server.

In the month following these failed Kerberos login attempts, between November 26 and December 22, Darktrace observed multiple internal devices encrypting files within the customer’s environment with the extensions “.cts1” and “.cts7”. Devices were also seen writing ransom notes with the file name “cAcTuS.readme.txt” to two additional internal devices, as well as files likely associated with Qlik, such as “QlikSense.pdf”. This activity detected by Darktrace confirmed the presence of a CACTUS ransomware infection that was spreading across the customer’s network.

The model, 'Ransom or Offensive Words Written to SMB', triggered in response to SMB file writes of the ransom note, ‘cAcTuS.readme.txt’, that was observed on the customer’s network.
Figure 5: The model, 'Ransom or Offensive Words Written to SMB', triggered in response to SMB file writes of the ransom note, ‘cAcTuS.readme.txt’, that was observed on the customer’s network.
CACTUS ransomware extensions, “.cts1” and “.cts7”, being appended to files on the customer’s network.
Figure 6: CACTUS ransomware extensions, “.cts1” and “.cts7”, being appended to files on the customer’s network.

Following this initial encryption activity, two affected devices were observed attempting to remove evidence of this activity by deleting the encrypted files.

Attackers attempting to remove evidence of their activity by deleting files with appendage “.cts1”.
Figure 7: Attackers attempting to remove evidence of their activity by deleting files with appendage “.cts1”.

Conclusion

In the face of this CACTUS ransomware attack, Darktrace’s anomaly-based approach to threat detection enabled it to quickly identify multiple stages of the cyber kill chain occurring in the customer’s environment. These stages ranged from ‘initial access’ by exploiting Qlik vulnerabilities, which Darktrace was able to detect before the method had been reported by external researchers, to ‘actions on objectives’ by encrypting files. Darktrace’s Self-Learning AI was also able to detect a previously unreported stage of the attack: multiple Kerberos brute force login attempts.

If Darktrace’s autonomous response capability, RESPOND, had been active and enabled in autonomous response mode at the time of this attack, it would have been able to take swift mitigative action to shut down such suspicious activity as soon as it was identified by DETECT, effectively containing the ransomware attack at the earliest possible stage.

Learning a network’s ‘normal’ to identify deviations from established patterns of behaviour enables Darktrace’s identify a potential compromise, even one that uses common and often legitimately used administrative tools. This allows Darktrace to stay one step ahead of the increasingly sophisticated TTPs used by ransomware actors.

Credit to Tiana Kelly, Cyber Analyst & Analyst Team Lead, Anna Gilbertson, Cyber Analyst

Appendices

References

[1] https://www.kroll.com/en/insights/publications/cyber/cactus-ransomware-prickly-new-variant-evades-detection

[2] https://www.bleepingcomputer.com/news/security/cactus-ransomware-exploiting-qlik-sense-flaws-to-breach-networks/

[3] https://explore.avertium.com/resource/new-ransomware-strains-cactus-and-3am

[4] https://www.soitron.com/cyber-attackers-abuse-holidays/

[5] https://arcticwolf.com/resources/blog/qlik-sense-exploited-in-cactus-ransomware-campaign/

Darktrace DETECT Models

Compromise / Agent Beacon (Long Period)

Anomalous Connection / PowerShell to Rare External

Device / New PowerShell User Agent

Device / Suspicious SMB Scanning Activity

Anomalous File / EXE from Rare External Location

Anomalous Connection / Unusual Internal Remote Desktop

User / Kerberos Password Brute Force

Compromise / Ransomware / Ransom or Offensive Words Written to SMB

Unusual Activity / Anomalous SMB Delete Volume

Anomalous Connection / Multiple Connections to New External TCP Port

Compromise / Slow Beaconing Activity To External Rare  

Compromise / SSL Beaconing to Rare Destination  

Anomalous Server Activity / Rare External from Server  

Compliance / Remote Management Tool On Server

Compromise / Agent Beacon (Long Period)  

Compromise / Suspicious File and C2  

Device / Internet Facing Device with High Priority Alert  

Device / Large Number of Model Breaches  

Anomalous File / Masqueraded File Transfer

Anomalous File / Internet facing System File Download  

Anomalous Server Activity / Outgoing from Server

Device / Initial Breach Chain Compromise  

Compromise / Agent Beacon (Medium Period)  

Compromise / Agent Beacon (Long Period)  

List of IoCs

IoC - Type - Description

zohoservice[.]net: 45.61.147[.]176 - Domain name: IP Address - Hosting payload over HTTP

Mozilla/5.0 (Windows NT; Windows NT 10.0; en-US) WindowsPowerShell/5.1.17763.2183 - User agent -PowerShell user agent

.cts1 - File extension - Malicious appendage

.cts7- File extension - Malicious appendage

cAcTuS.readme.txt - Filename -Ransom note

putty.zip – Filename - Initial payload: ZIP containing PuTTY Link

MITRE ATT&CK Mapping

Tactic - Technique  - SubTechnique

Web Protocols: COMMAND AND CONTROL - T1071 -T1071.001

Powershell: EXECUTION - T1059 - T1059.001

Exploitation of Remote Services: LATERAL MOVEMENT - T1210 – N/A

Vulnerability Scanning: RECONAISSANCE     - T1595 - T1595.002

Network Service Scanning: DISCOVERY - T1046 - N/A

Malware: RESOURCE DEVELOPMENT - T1588 - T1588.001

Drive-by Compromise: INITIAL ACCESS - T1189 - N/A

Remote Desktop Protocol: LATERAL MOVEMENT – 1021 -T1021.001

Brute Force: CREDENTIAL ACCESS        T – 1110 - N/A

Data Encrypted for Impact: IMPACT - T1486 - N/A

Data Destruction: IMPACT - T1485 - N/A

File Deletion: DEFENSE EVASION - T1070 - T1070.004

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
Tiana Kelly
Senior Cyber Analyst & Team Lead

More in this series

No items found.

Blog

/

AI

/

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

/

AI

/

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
あなたのデータ × DarktraceのAI
唯一無二のDarktrace AIで、ネットワークセキュリティを次の次元へ