Blog
/
Network
/
February 23, 2024

Quasar Remote Access Tool and Its Security Risks

Discover how the Quasar remote access tool can become a vulnerability in the wrong hands and strategies to mitigate these risks.
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
Nicole Wong
Cyber Security Analyst
Default blog image
23
Feb 2024

The threat of interoperability

As the “as-a-Service” market continues to grow, indicators of compromise (IoCs) and malicious infrastructure are often interchanged and shared between multiple malware strains and attackers. This presents organizations and their security teams with a new threat: interoperability.

Interoperable threats not only enable malicious actors to achieve their objectives more easily by leveraging existing infrastructure and tools to launch new attacks, but the lack of clear attribution often complicates identification for security teams and incident responders, making it challenging to mitigate and contain the threat.

One such threat observed across the Darktrace customer base in late 2023 was Quasar, a legitimate remote administration tool that has becoming increasingly popular for opportunistic attackers in recent years. Working in tandem, the anomaly-based detection of Darktrace DETECT™ and the autonomous response capabilities of Darktrace RESPOND™ ensured that affected customers were promptly made aware of any suspicious activity on the attacks were contained at the earliest possible stage.

What is Quasar?

Quasar is an open-source remote administration tool designed for legitimate use; however, it has evolved to become a popular tool used by threat actors due to its wide array of capabilities.  

How does Quasar work?

For instance, Quasar can perform keylogging, take screenshots, establish a reverse proxy, and download and upload files on a target device [1].  A report released towards the end of 2023 put Quasar back on threat researchers’ radars as it disclosed the new observation of dynamic-link library (DLL) sideloading being used by malicious versions of this tool to evade detection [1].  DLL sideloading involves configuring legitimate Windows software to run a malicious file rather than the legitimate file it usually calls on as the software loads.  The evolving techniques employed by threat actors using Quasar highlights defenders’ need for anomaly-based detections that do not rely on pre-existing knowledge of attacker techniques, and can identify and alert for unusual behavior, even if it is performed by a legitimate application.

Although Quasar has been used by advanced persistent threat (APT) groups for global espionage operations [2], Darktrace observed the common usage of default configurations for Quasar, which appeared to use shared malicious infrastructure, and occurred alongside other non-compliant activity such as BitTorrent use and cryptocurrency mining.  

Quasar Attack Overview and Darktrace Coverage

Between September and October 2023, Darktrace detected multiple cases of malicious Quasar activity across several customers, suggesting probable campaign activity.  

Quasar infections can be difficult to detect using traditional network or host-based tools due to the use of stealthy techniques such as DLL side-loading and encrypted SSL connections for command-and control (C2) communication, that traditional security tools may not be able to identify.  The wide array of capabilities Quasar possesses also suggests that attacks using this tool may not necessarily be modelled against a linear kill chain. Despite this, the anomaly-based detection of Darktrace DETECT allowed it to identify IoCs related to Quasar at multiple stages of the kill chain.

Quasar Initial Infection

During the initial infection stage of a Quasar compromise observed on the network of one customer, Darktrace detected a device downloading several suspicious DLL and executable (.exe) files from multiple rare external sources using the Xmlst user agent, including the executable ‘Eppzjtedzmk[.]exe’.  Analyzing this file using open-source intelligence (OSINT) suggests this is a Quasar payload, potentially indicating this represented the initial infection through DLL sideloading [3].

Interestingly, the Xmlst user agent used to download the Quasar payload has also been associated with Raccoon Stealer, an information-stealing malware that also acts as a dropper for other malware strains [4][5]. The co-occurrence of different malware components is increasingly common across the threat landscape as MaaS operating models increases in popularity, allowing attackers to employ cross-functional components from different strains.

Figure 1: Cyber AI Analyst Incident summarizing the multiple different downloads in one related incident, with technical details for the Quasar payload included. The incident event for Suspicious File Download is also linked to Possible HTTP Command and Control, suggesting escalation of activity following the initial infection.  

Quasar Establishing C2 Communication

During this phase, devices on multiple customer networks were identified making unusual external connections to the IP 193.142.146[.]212, which was not commonly seen in their networks. Darktrace analyzed the meta-properties of these SSL connections without needing to decrypt the content, to alert the usage of an unusual port not typically associated with the SSL protocol, 4782, and the usage of self-signed certificates.  Self-signed certificates do not provide any trust value and are commonly used in malware communications and ill-reputed web servers.  

Further analysis into these alerts using OSINT indicated that 193.142.146[.]212 is a Quasar C2 server and 4782 is the default port used by Quasar [6][7].  Expanding on the self-signed certificate within the Darktrace UI (see Figure 3) reveals a certificate subject and issuer of “CN=Quasar Server CA”, which is also the default self-signed certificate compiled by Quasar [6].

Figure 2: Cyber AI Analyst Incident summarizing the repeated external connections to a rare external IP that was later associated with Quasar.
Figure 3: Device Event Log of the affected device, showing Darktrace’s analysis of the SSL Certificate associated with SSL connections to 193.142.146[.]212.

A number of insights can be drawn from analysis of the Quasar C2 endpoints detected by Darktrace across multiple affected networks, suggesting a level of interoperability in the tooling used by different threat actors. In one instance, Darktrace detected a device beaconing to the endpoint ‘bittorrents[.]duckdns[.]org’ using the aforementioned “CN=Quasar Server CA” certificate. DuckDNS is a dynamic DNS service that could be abused by attackers to redirect users from their intended endpoint to malicious infrastructure, and may be shared or reused in multiple different attacks.

Figure 4: A device’s Model Event Log, showing the Quasar Server CA SSL certificate used in connections to 41.233.139[.]145 on port 5, which resolves via passive replication to ‘bittorrents[.]duckdns[.]org’.  

The sharing of malicious infrastructure among threat actors is also evident as several OSINT sources have also associated the Quasar IP 193.142.146[.]212, detected in this campaign, with different threat types.

While 193.142.146[.]212:4782 is known to be associated with Quasar, 193.142.146[.]212:8808 and 193.142.146[.]212:6606 have been associated with AsyncRAT [11], and the same IP on port 8848 has been associated with RedLineStealer [12].  Aside from the relative ease of using already developed tooling, threat actors may prefer to use open-source malware in order to avoid attribution, making the true identity of the threat actor unclear to incident responders [1][13].  

Quasar Executing Objectives

On multiple customer deployments affected by Quasar, Darktrace detected devices using BitTorrent and performing cryptocurrency mining. While these non-compliant, and potentially malicious, activities are not necessarily specific IoCs for Quasar, they do suggest that affected devices may have had greater attack surfaces than others.

For instance, one affected device was observed initiating connections to 162.19.139[.]184, a known Minergate cryptomining endpoint, and ‘zayprostofyrim[.]zapto[.]org’, a dynamic DNS endpoint linked to the Quasar Botnet by multiple OSINT vendors [9].

Figure 5: A Darktrace DETECT Event Log showing simultaneous connections to a Quasar endpoint and a cryptomining endpoint 162.19.139[.]184.

Not only does cryptocurrency mining use a significant amount of processing power, potentially disrupting an organization’s business operations and racking up high energy bills, but the software used for this mining is often written to a poor standard, thus increasing the attack surfaces of devices using them. In this instance, Quasar may have been introduced as a secondary payload from a user or attacker-initiated download of cryptocurrency mining malware.

Similarly, it is not uncommon for malicious actors to attach malware to torrented files and there were a number of examples of Darktrace detect identifying non-compliant activity, like BitTorrent connections, overlapping with connections to external locations associated with Quasar. It is therefore important for organizations to establish and enforce technical and policy controls for acceptable use on corporate devices, particularly when remote working introduces new risks.  

Figure 6: A device’s Event Log filtered by Model Breaches, showing a device connecting to BitTorrent shortly before making new or repeated connections to unusual endpoints, which were subsequently associated to Quasar.

In some cases observed by Darktrace, devices affected by Quasar were also being used to perform data exfiltration. Analysis of a period of unusual external connections to the aforementioned Quasar C2 botnet server, ‘zayprostofyrim[.]zapto[.]org’, revealed a small data upload, which may have represented the exfiltration of some data to attacker infrastructure.

Darktrace’s Autonomous Response to Quasar Attacks

On customer networks that had Darktrace RESPOND™ enabled in autonomous response mode, the threat of Quasar was mitigated and contained as soon as it was identified by DETECT. If RESPOND is not configured to respond autonomously, these actions would instead be advisory, pending manual application by the customer’s security team.

For example, following the detection of devices downloading malicious DLL and executable files, Darktrace RESPOND advised the customer to block specific connections to the relevant IP addresses and ports. However, as the device was seen attempting to download further files from other locations, RESPOND also suggested enforced a ‘pattern of life’ on the device, meaning it was only permitted to make connections that were part its normal behavior. By imposing a pattern of life, Darktrace RESPOND ensures that a device cannot perform suspicious behavior, while not disrupting any legitimate business activity.

Had RESPOND been configured to act autonomously, these mitigative actions would have been applied without any input from the customer’s security team and the Quasar compromise would have been contained in the first instance.

Figure 7: The advisory actions Darktrace RESPOND initiated to block specific connections to a malicious IP and to enforce the device’s normal patterns of life in response to the different anomalies detected on the device.

In another case, one customer affected by Quasar did have enabled RESPOND to take autonomous action, whilst also integrating it with a firewall. Here, following the detection of a device connecting to a known Quasar IP address, RESPOND initially blocked it from making connections to the IP via the customer’s firewall. However, as the device continued to perform suspicious activity after this, RESPOND escalated its response by blocking all outgoing connections from the device, effectively preventing any C2 activity or downloads.

Figure 8: RESPOND actions triggered to action via integrated firewall and TCP Resets.

Conclusion

When faced with a threat like Quasar that utilizes the infrastructure and tools of both legitimate services and other malicious malware variants, it is essential for security teams to move beyond relying on existing knowledge of attack techniques when safeguarding their network. It is no longer enough for organizations to rely on past attacks to defend against the attacks of tomorrow.

Crucially, Darktrace’s unique approach to threat detection focusses on the anomaly, rather than relying on a static list of IoCs or "known bads” based on outdated threat intelligence. In the case of Quasar, alternative or future strains of the malware that utilize different IoCs and TTPs would still be identified by Darktrace as anomalous and immediately alerted.

By learning the ‘normal’ for devices on a customer’s network, Darktrace DETECT can recognize the subtle deviations in a device’s behavior that could indicate an ongoing compromise. Darktrace RESPOND is subsequently able to follow this up with swift and targeted actions to contain the attack and prevent it from escalating further.

Credit to Nicole Wong, Cyber Analyst, Vivek Rajan Cyber Analyst

Appendices

Darktrace DETECT Model Breaches

  • Anomalous Connection / Multiple Failed Connections to Rare Endpoint
  • Anomalous Connection / Anomalous SSL without SNI to New External
  • Anomalous Connection / Application Protocol on Uncommon Port
  • Anomalous Connection / Rare External SSL Self-Signed
  • Compromise / New or Repeated to Unusual SSL Port
  • Compromise / Beaconing Activity To External Rare
  • Compromise / High Volume of Connections with Beacon Score
  • Compromise / Large Number of Suspicious Failed Connections
  • Unusual Activity / Unusual External Activity

List of IoCs

IP:Port

193.142.146[.]212:4782 -Quasar C2 IP and default port

77.34.128[.]25: 8080 - Quasar C2 IP

Domain

zayprostofyrim[.]zapto[.]org - Quasar C2 Botnet Endpoint

bittorrents[.]duckdns[.]org - Possible Quasar C2 endpoint

Certificate

CN=Quasar Server CA - Default certificate used by Quasar

Executable

Eppzjtedzmk[.]exe - Quasar executable

IP Address

95.214.24[.]244 - Quasar C2 IP

162.19.139[.]184 - Cryptocurrency Miner IP

41.233.139[.]145[VR1] [NW2] - Possible Quasar C2 IP

MITRE ATT&CK Mapping

Command and Control

T1090.002: External Proxy

T1071.001: Web Protocols

T1571: Non-Standard Port

T1001: Data Obfuscation

T1573: Encrypted Channel

T1071: Application Layer Protocol

Resource Development

T1584: Compromise Infrastructure

References

[1] https://thehackernews.com/2023/10/quasar-rat-leverages-dll-side-loading.html

[2] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/cicada-apt10-japan-espionage

[3]https://www.virustotal.com/gui/file/bd275a1f97d1691e394d81dd402c11aaa88cc8e723df7a6aaf57791fa6a6cdfa/community

[4] https://twitter.com/g0njxa/status/1691826188581298389

[5] https://www.linkedin.com/posts/grjk83_raccoon-stealer-announce-return-after-hiatus-activity-7097906612580802560-1aj9

[6] https://community.netwitness.com/t5/netwitness-community-blog/using-rsa-netwitness-to-detect-quasarrat/ba-p/518952

[7] https://www.cisa.gov/news-events/analysis-reports/ar18-352a

[8]https://any.run/report/6cf1314c130a41c977aafce4585a144762d3fb65f8fe493e836796b989b002cb/7ac94b56-7551-4434-8e4f-c928c57327ff

[9] https://threatfox.abuse.ch/ioc/891454/

[10] https://www.virustotal.com/gui/ip-address/41.233.139.145/relations

[11] https://raw.githubusercontent.com/stamparm/maltrail/master/trails/static/malware/asyncrat.txt

[12] https://sslbl.abuse.ch/ssl-certificates/signature/RedLineStealer/

[13] https://www.botconf.eu/botconf-presentation-or-article/hunting-the-quasar-family-how-to-hunt-a-malware-family/

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
Nicole Wong
Cyber Security Analyst

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