Blog
/
Network
/
March 19, 2024

Pikabot Malware: Insights, Impact, & Attack Analysis

Learn about Pikabot malware and its rapid evolution in the wild, impacting organizations and how to defend against this growing threat.
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
Brianna Luong (Leddy)
Sr. Technical Alliances Manager
Default blog image
19
Mar 2024

How does Loader Malware work?

Throughout 2023, the Darktrace Threat Research team identified and investigated multiple strains of loader malware affecting customers across its fleet. These malicious programs typically serve as a gateway for threat actors to gain initial access to an organization’s network, paving the way for subsequent attacks, including additional malware infections or disruptive ransomware attacks.

How to defend against loader malware

The prevalence of such initial access threats highlights the need for organizations to defend against multi-phase compromises, where modular malware swiftly progresses from one stage of an attack to the next. One notable example observed in 2023 was Pikabot, a versatile loader malware used for initial access and often accompanied by secondary compromises like Cobalt Strike and Black Basta ransomware.

While Darktrace initially investigated multiple instances of campaign-like activity associated with Pikabot during the summer of 2023, a new campaign emerged in October which was observed targeting a Darktrace customer in Europe. Thanks to the timely detection by Darktrace DETECT™ and the support of Darktrace’s Security Operations Center (SOC), the Pikabot compromise was quickly shut down before it could escalate into a more disruptive attack.

What is Pikabot?

Pikabot is one of the latest modular loader malware strains that has been active since the first half of 2023, with several evolutions in its methodology observed in the months since. Initial researchers noted similarities to the Qakbot aka Qbot or Pinkslipbot and Mantanbuchus malware families, and while Pikabot appears to be a new malware in early development, it shares multiple commonalities with Qakbot [1].

First, both Pikabot and Qakbot have similar distribution methods, can be used for multi-stage attacks, and are often accompanied by downloads of Cobalt Strike and other malware strains. The threat actor known as TA577, which has also been referred to as Water Curupira, has been seen to use both types of malware in spam campaigns which can lead to Black Basta ransomware attacks [2] [3].Notably, a rise in Pikabot campaigns were observed in September and October 2023, shortly after the takedown of Qakbot in Operation Duck Hunt, suggesting that Pikabot may be serving as a replacement for initial access to target network [4].

How does Pikabot malware work?

Many Pikabot infections start with a malicious email, particularly using email thread hijacking; however, other cases have been distributed via malspam and malvertising [5]. Once downloaded, Pikabot runs anti-analysis techniques and checks the system’s language, self-terminating if the language matches that of a Commonwealth of Independent States (CIS) country, such as Russian or Ukrainian. It will then gather key information to send to a command-and-control (C2) server, at which point additional payload downloads may be observed [2]. Early response to a Pikabot infection is important for organizations to prevent escalation to a significant compromise such as ransomware.

Darktrace’s Coverage of Pikabot malware

Between April and July 2023, the Darktrace Threat Research team investigated Pikabot infections affected more than 15 customer environments; these attacks primarily targeted US and European organizations spanning multiple industries, and most followed the below lifecycle:

  1. Initial access via malspam or email, often outside of Darktrace’s scope
  2. Suspicious executable download from a URI in the format /\/[a-z0-9A-Z]{3,}\/[a-z0-9A-Z]{5,}/ and using a Windows PowerShell user agent
  3. C2 connections to IP addresses on uncommon ports including 1194 and 2078
  4. Some cases involved further C2 activity to Cobalt Strike endpoints

In October 2023, a second campaign emerged that largely followed the same attack pattern, with a notable difference that cURL was used for the initial payload download as opposed to PowerShell. All the Pikabot cases that Darktrace has observed since October 2023 have used cURL, which could indicate a shift in approach from targeting Windows devices to multi-operating system environments.

Figure 1: Timeline of the Pikabot infection over a 2-hour period.

On October 17, 2023, Darktrace observed a Pikabot infection on the network of a European customer after an internal user seemingly clicked a malicious link in a phishing email, thereby compromising their device. As the customer did not have Darktrace/Email™ deployed on their network, Darktrace did not have visibility over the email. Despite this, DETECT was still able to provide full visibility over the network-based activity that ensued.

Darktrace observed the device using a cURL user agent when initiating the download of an unusual executable (.exe) file from an IP address that had never previously been observed on the network. Darktrace further recognized that the executable file was attempting to masquerade as a different file type, likely to evade the detection of security teams and their security tools. Within one minute, the device began to communicate with additional unusual IP addresses on uncommon ports (185.106.94[.]174:5000 and 80.85.140[.]152:5938), both of which have been noted by open-source intelligence (OSINT) vendors as Pikabot C2 servers [6] [7].

Figure 2: Darktrace model breach Event Log showing the initial file download, immediately followed by a connection attempt to a Pikabot C2 server.

Around 40 minutes after the initial download, Darktrace detected the device performing suspicious DNS tunneling using a pattern that resembled the Cobalt Strike Beacon. This was accompanied by beaconing activity to a rare domain, ‘wordstt182[.]com’, which was registered only 4 days prior to this activity [8]. Darktrace observed additional DNS connections to the endpoint, ‘building4business[.]net’, which had been linked to Black Basta ransomware [2].

Figure 3: The affected device making successful TXT DNS requests to known Black Basta endpoints.

As this customer had integrated Darktrace with the Microsoft Defender, Defender was able to contextualize the DETECT model breaches with endpoint insights, such as known threats and malware, providing customers with unparalleled visibility of the host-level detections surrounding network-level anomalies.

In this case, the behavior of the affected device triggered multiple Microsoft Defender alerts, including one alert which linked the activity to the threat actor Storm-0464, another name for TA577 and Water Curupira. These insights were presented to the customer in the form of a Security Integration alert, allowing them to build a full picture of the ongoing incident.

Figure 4: Security Integration alert from Microsoft Defender in Darktrace, linking the observed activity to the threat group Storm-0464.

As the customer had subscribed to Darktrace’s Proactive Threat Notification (PTN) service, the customer received timely alerts from Darktrace’s SOC notifying them of the suspicious activity associated with Pikabot. This allowed the customer’s security team to quickly identify the affected device and remove it from their environment for remediation.

Although the customer did have Darktrace RESPOND™ enabled on their network, it was configured in human confirmation mode, requiring manual application for any RESPOND actions. RESPOND had suggested numerous actions to interrupt and contain the attack, including blocking connections to the observed Pikabot C2 addresses, which were manually actioned by the customer’s security team after the fact. Had RESPOND been enabled in autonomous response mode during the attack, it would have autonomously blocked these C2 connections and prevented the download of any suspicious files, effectively halting the escalation of the attack.

Nonetheless, Darktrace DETECT’s prompt identification and alerting of this incident played a crucial role in enabling the customer to mitigate the threat of Pikabot, preventing it from progressing into a disruptive ransomware attack.

Figure 5: Darktrace RESPOND actions recommended from the initial file download and throughout the C2 traffic, ranging from blocking specific connections to IP addresses and ports to enforcing a normal pattern of life for the source device.

Conclusion

Pikabot is just one recent example of a modular strain of loader known for its adaptability and speed, seamlessly changing tactics from one campaign to the next and utilizing new infrastructure to initiate multi-stage attacks. Leveraging commonly used tools and services like Windows PowerShell and cURL, alongside anti-analysis techniques, this malware can evade the detection and often bypass traditional security tools.

In this incident, Darktrace detected a Pikabot infection in its early stages, identifying an anomalous file download using a cURL user agent, a new tactic for this particular strain of malware. This timely detection, coupled with the support of Darktrace’s SOC, empowered the customer to quickly identify the compromised device and act against it, thwarting threat actors attempting to connect to malicious Cobalt Strike and Black Basta servers. By preventing the escalation of the attack, including potential ransomware deployment, the customer’s environment remained safeguarded.

Had Darktrace RESPOND been enabled in autonomous response mode at the time of this attack, it would have been able to further support the customer by applying targeted mitigative actions to contain the threat of Pikabot at its onset, bolstering their defenses even more effectively.

Credit to Brianna Leddy, Director of Analysis, Signe Zaharka, Senior Cyber Security Analyst

Appendix

Darktrace DETECT Models

Anomalous Connection / Anomalous SSL without SNI to New External

Anomalous Connection / Application Protocol on Uncommon Port

Anomalous Connection / Multiple Connections to New External TCP Port

Anomalous Connection / New User Agent to IP Without Hostname

Anomalous Connection / Powershell to Rare External

Anomalous Connection / Rare External SSL Self-Signed

Anomalous Connection / Repeated Rare External SSL Self-Signed

Anomalous File / EXE from Rare External Location

Anomalous File / Masqueraded File Transfer

Anomalous File / Multiple EXE from Rare External Locations

Compromise / Agent Beacon to New Endpoint

Compromise / Beacon to Young Endpoint

Compromise / Beaconing Activity To External Rare

Compromise / DNS / DNS Tunnel with TXT Records

Compromise / New or Repeated to Unusual SSL Port

Compromise / SSL Beaconing to Rare Destination

Compromise / Suspicious Beaconing Behaviour

Compromise / Suspicious File and C2

Device / Initial Breach Chain Compromise

Device / Large Number of Model Breaches

Device / New PowerShell User Agent

Device / New User Agent

Device / New User Agent and New IP

Device / Suspicious Domain

Security Integration / C2 Activity and Integration Detection

Security Integration / Egress and Integration Detection

Security Integration / High Severity Integration Detection

Security Integration / High Severity Integration Incident

Security Integration / Low Severity Integration Detection

Security Integration / Low Severity Integration Incident

Antigena / Network / External Threat / Antigena File then New Outbound Block

Antigena / Network / External Threat / Antigena Suspicious Activity Block

Antigena / Network / External Threat / Antigena Suspicious File Block

Antigena / Network / Significant Anomaly / Antigena Breaches Over Time Block

Antigena / Network / Significant Anomaly / Antigena Controlled and Model Breach

Antigena / Network / Significant Anomaly / Antigena Enhanced Monitoring from Client Block

Antigena / Network / Significant Anomaly / Antigena Significant Anomaly from Client Block

Antigena / Network / Significant Anomaly / Antigena Significant Security Integration and Network Activity Block

List of Indicators of Compromise (IoC)

IOC - TYPE - DESCRIPTION + CONFIDENCE

128.140.102[.]132 - IP Address - Pikabot Download

185.106.94[.]174:5000 - IP Address: Port - Pikabot C2 Endpoint

80.85.140[.]152:5938 - IP Address: Port - Pikabot C2 Endpoint

building4business[.]net - Hostname - Cobalt Strike DNS Beacon

wordstt182[.]com - Hostname - Cobalt Strike Server

167.88.166[.]109 - IP Address - Cobalt Strike Server

192.9.135[.]73 - IP - Pikabot C2 Endpoint

192.121.17[.]68 - IP - Pikabot C2 Endpoint

185.87.148[.]132 - IP - Pikabot C2 Endpoint

129.153.22[.]231 - IP - Pikabot C2 Endpoint

129.153.135[.]83 - IP - Pikabot C2 Endpoint

154.80.229[.]76 - IP - Pikabot C2 Endpoint

192.121.17[.]14 - IP - Pikabot C2 Endpoint

162.252.172[.]253 - IP - Pikabot C2 Endpoint

103.124.105[.]147 - IP - Likely Pikabot Download

178.18.246[.]136 - IP - Pikabot C2 Endpoint

86.38.225[.]106 - IP - Pikabot C2 Endpoint

198.44.187[.]12 - IP - Pikabot C2 Endpoint

154.12.233[.]66 - IP - Pikabot C2 Endpoint

MITRE ATT&CK Mapping

TACTIC - TECHNIQUE

Defense Evasion - Masquerading: Masquerade File Type (T1036.008)

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

Command and Control - Non-Standard Port (T1571)

Command and Control - Application Layer Protocol: DNS (T1071.004)

Command and Control - Protocol Tunneling (T1572)

References

[1] https://news.sophos.com/en-us/2023/06/12/deep-dive-into-the-pikabot-cyber-threat/?&web_view=true  

[2] https://www.trendmicro.com/en_be/research/24/a/a-look-into-pikabot-spam-wave-campaign.html

[3] https://thehackernews.com/2024/01/alert-water-curupira-hackers-actively.html

[4] https://www.darkreading.com/cyberattacks-data-breaches/pikabot-malware-qakbot-replacement-black-basta-attacks

[5] https://www.redpacketsecurity.com/pikabot-distributed-via-malicious-ads-6/

[6] https://www.virustotal.com/gui/ip-address/185.106.94.174/detection

[7] https://www.virustotal.com/gui/ip-address/80.85.140.152/detection

[8] https://www.domainiq.com/domain?wordstt182.com

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
Brianna Luong (Leddy)
Sr. Technical Alliances Manager

More in this series

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