Blog
/
Network
/
March 22, 2024

What are Botnets and How Darktrace Uncovers Them

Learn how Darktrace detected and implemented defense protocols against Socks5Systemz botnet before any threat to intelligence had been published.
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
Adam Potter
Senior Cyber Analyst
Default blog image
22
Mar 2024

What are botnets?

Although not a recent addition to the threat landscape, botnets persist as a significant concern for organizations, with many threat actors utilizing them for political, strategic, or financial gain. Botnets pose a particularly persistent threat to security teams; even if one compromised device is detected, attackers will likely have infected multiple devices and can continue to operate. Moreover, threat actors are able to easily replace the malware communication channels between infected devices and their command-and-control (C2) servers, making it incredibly difficult to remove the infection.

Botnet example: Socks5Systemz

One example of a botnet recently investigated by the Darktrace Threat Research team is Socks5Systemz. Socks5Systemz is a proxy-for-rent botnet, whereby actors can rent blocks of infected devices to perform proxying services.  Between August and November 2023, Darktrace detected indicators of Socks5Systemz botnet compromise within a cross-industry section of the customer base. Although open-source intelligence (OSINT) research of the botnet only appeared in November 2023, the anomaly-based approach of Darktrace DETECT™ allowed it to identify multiple stages of the network-based activity on affected customer systems well before traditional rules and signatures would have been implemented.

Darktrace’s Cyber AI Analyst™ complemented DETECT’s successful identification of Socks5Systemz activity on customer networks, playing a pivotal role in piecing together the seemingly separate events that comprised the wider compromise. This allowed Darktrace to build a clearer picture of the attack, empowering its customers with full visibility over emerging incidents.

In the customer environments highlighted in this blog, Darktrace RESPOND™ was not configured to operate autonomously. As a result, Socks5Systemz attacks were able to advance through their kill chains until customer security teams acted upon Darktrace’s detections and began their remediation procedures.

What is Socks5Systemz?

The Socks5Systemz botnet is a proxy service where individuals can use infected devices as proxy servers.

These devices act as ‘middlemen’, forwarding connections from malicious actors on to their intended destination. As this additional connectivity conceals the true origin of the connections, threat actors often use botnets to increase their anonymity. Although unauthorized proxy servers on a corporate network may not appear at first glance to be a priority for organizations and their security teams, complicity in proxy botnets could result in reputational damage and significant financial losses.

Since it was first observed in the wild in 2016, the Socks5Systemz botnet has grown steadily, seemingly unnoticed by cyber security professionals, and has infected a reported 10,000 devices worldwide [1]. Cyber security researchers noted a high concentration of compromised devices in India, with lower concentrations of devices infected in the United States, Latin America, Australia and multiple European and African countries [2]. Renting sections of the Socks5Systemz botnet costs between 1 USD and 4,000 USD, with options to increase the threading and time-range of the rentals [2]. Due to the lack of affected devices in Russia, some threat researchers have concluded that the botnet’s operators are likely Russian [2].

Darktrace’s Coverage of Socks5Systemz

The Darktrace Threat Research team conducted investigations into campaign-like activity across the customer base between August and November 2023, where multiple indicators of compromise (IoCs) relating to the Socks5Systemz proxy botnet were observed. Darktrace identified several stages of the attack chain described in static malware analysis by external researchers. Darktrace was also able to uncover additional IoCs and stages of the Socks5Systemz attack chain that had not featured in external threat research.

Delivery and Execution

Prior research on Socks5Systemz notes how the malware is typically delivered via user input, with delivery methods including phishing emails, exploit kits, malicious ads, and trojanized executables downloaded from peer-to-peer (P2P) networks [1].

Threat actors have also used separate malware loaders such as PrivateLoader and Amadey deliver the Socks5Systemz payload. These loaders will drop executable files that are responsible for setting up persistence and injecting the proxy bot into the infected device’s memory [2]. Although evidence of initial payload delivery did not appear during its investigations, Darktrace did discover IoCs relating to PrivateLoader and Amadey on multiple customer networks. Such activity included HTTP POST requests using PHP to rare external IPs and HTTP connections with a referrer header field, indicative of a redirected connection.

However, additional adjacent activity that may suggest initial user execution and was observed during Darktrace’s investigations. For example, an infected device on one deployment made a HTTP GET request to a rare external domain with a “.fun” top-level domain (TLD) for a PDF file. The URI also appears to have contained a client ID. While this download and HTTP request likely corresponded to the gathering and transmission of further telemetry data and infection verification [2], the downloaded PDF file may have represented a malicious payload.

Advanced Search log details highlighting a device infected by Socks5Systemz downloading a suspicious PDF file.
Figure 1: Advanced Search log details highlighting a device infected by Socks5Systemz downloading a suspicious PDF file.

Establishing C2 Communication  

Once the proxy bot has been injected into the device’s memory, the malware attempts to contact servers owned by the botnet’s operators. Across several customer environments, Darktrace identified infected devices attempting to establish connections with such C2 servers. First, affected devices would make repeated HTTP GET requests over port 80 to rare external domains; these endpoints typically had “.ua” and “.ru” TLDs. The majority of these connection attempts were not preceded by a DNS host lookup, suggesting that the domains were already loaded in the device’s cache memory or hardcoded into the code of running processes.

Figure 2: Breach log data connections identifying repeated unusual HTTP connections over port 80 for domains without prior DNS host lookup.

While most initial HTTP GET requests across investigated incidents did not feature DNS host lookups, Darktrace did identify affected devices on a small number of customer environments performing a series of DNS host lookups for seemingly algorithmically generated domains (DGA). These domains feature the same TLDs as those seen in connections without prior DNS host lookups.  

Figure 3: Cyber AI Analyst data indicating a subset of DGAs queried via DNS by infected devices.

These DNS requests follow the activity reported by researchers, where infected devices query a hardcoded DNS server controlled by the threat actor for an DGA domain [2]. However, as the bulk of Darktrace’s investigations presented HTTP requests without a prior DNS host lookup, this activity indicates a significant deviation from the behavior reported by OSINT sources. This could indicate that multiple variations of the Socks5Systemz botnet were circulating at the time of investigation.

Most hostnames observed during this time of investigation follow a specific regular expression format: /[a-z]{7}\.(ua|net|info|com|ru)/ or /[a-z0-9]{15}\.(ua)/. Darktrace also noticed the HTTP GET requests for DGA domains followed a consistent URI pattern: /single.php?c=<STRING>. The requests were also commonly made using the “Mozilla/5.0 (Windows; U; MSIE 9.0; Windows NT 9.0; en-US)” user agent over port 80.

This URI pattern observed during Darktrace’s investigations appears to reflect infected devices contacting Socks5Systemz C2 servers to register the system and details of the host, and signal it is ready to receive further instructions [2]. These URIs are encrypted with a RC4 stream cipher and contain information relating to the device’s operating system and architecture, as well as details of the infection.

The HTTP GET requests during this time, which involved devices made to a variety a variety of similar DGA domains, appeared alongside IP addresses that were later identified as Socks5Systemz C2 servers.

Figure 4: Cyber AI Analyst investigation details highlighting HTTP GET activity whereby RC4 encrypted data is sent to proxy C2 domains.

However, not all affected devices observed by Darktrace used DGA domains to transmit RC4 encoded data. Some investigated systems were observed making similar HTTP GET requests over port 80, albeit to the external domain: “bddns[.]cc”, using the aforementioned Mozilla user agent. During these requests, Darktrace identified a consistent URI pattern, similar to that seen in the DGA domain GET requests: /sign/<RC4 cipher text>.  

Darktrace DETECT recognized the rarity of the domains and IPs that were connected to by affected devices, as well as the usage of the new Mozilla user agent.  The HTTP connections, and the corresponding Darktrace DETECT model breaches, parallel the analysis made by external researchers: if the initial DGA DNS requests do not return a valid C2 server, infected devices connect to, and request the IP address of a server from, the above-mentioned domain [2].

Connection to Proxy

After sending host and infection details via HTTP and receiving commands from the C2 server, affected devices were frequently observed initiating activity to join the Sock5Systemz botnet. Infected hosts would first make HTTP GET requests to an IP identified as Socks5Systemz’s proxy checker application, usually sending the URI “proxy-activity.txt” to the domain over the HTTP protocol. This likely represents an additional validation check to confirm that the infected device is ready to join the botnet.

Figure 5: Cyber AI Analyst investigation detailing HTTP GET requests over port 80 to the Socks5Systemz Proxy Checker Application.

Following the final validation checks, devices would then attempt TCP connections to a range of IPs, which have been associated with BackConnect proxy servers, over port 1074. At this point, the device is able to receive commands from actors who login to and operate the corresponding BackConnect server. This BackConnect server will transmit traffic from the user renting the segment of the botnet [2].

Darktrace observed a range of activity associated with this stage of the attack, including the use of new or unusual user agents, connections to suspicious IPs, and other anomalous external connectivity which represented a deviation from affected devices’ expected behavior.

Additional Activities Following Proxy Addition

The Darktrace Threat Research team found evidence of the possible deployment of additional malware strains during their investigation into devices affected by Socks5Systemz. IoCs associated with both the Amadey and PrivateLoader loader malware strains, both of which are known to distribute Socks5Systemz, were also observed on affected devices. Additionally, Darktrace observed multiple infected systems performing cryptocurrency mining operations around the time of the Sock5Systemz compromise, utilizing the MinerGate protocol to conduct login and job functions, as well as making DNS requests for mining pools.

While such behavior would fall outside of the expected activity for Socks5Systemz and cannot be definitively attributed to it, Darktrace did observe devices affected by the botnet performing additional malicious downloads and operations during its investigations.

Conclusion

Ultimately, Darktrace’s anomaly-based approach to threat detection enabled it to effectively identify and alert for malicious Socks5Systemz botnet activity long before external researchers had documented its IoCs and tactics, techniques, and procedures (TTPs).  

In fact, Darktrace not only identified multiple distinct attack phases later outlined in external research but also uncovered deviations from these expected patterns of behavior. By proactively detecting emerging threats through anomaly detection rather than relying on existing threat intelligence, Darktrace is well positioned to detect evolving threats like Socks5Systemz, regardless of what their future iterations might look like.

Faced with the threat of persistent botnets, it is crucial for organizations to detect malicious activity in its early stages before additional devices are compromised, making it increasingly difficult to remediate. Darktrace’s suite of products enables the swift and effective detection of such threats. Moreover, when enabled in autonomous response mode, Darktrace RESPOND is uniquely positioned to take immediate, targeted actions to contain these attacks from the onset.

Credit to Adam Potter, Cyber Security Analyst, Anna Gilbertson, Cyber Security Analyst

Appendices

DETECT Model Breaches

  • Anomalous Connection / Multiple Failed Connections to Rare Endpoint
  • Anomalous Connection / Multiple Connections to New External TCP Port
  • Compromise / Beaconing Activity To External Rare
  • Compromise / DGA Beacon
  • Compromise / Beacon to Young Endpoint
  • Compromise / Slow Beaconing Activity To External Rare
  • Compromise / HTTP Beaconing to Rare Destination
  • Compromise / Quick and Regular Windows HTTP Beaconing
  • Compromise / Agent Beacon (Medium Period)
  • Compromise / Agent Beacon (Long Period)
  • Device / New User Agent
  • Device / New User Agent and New IP

Cyber AI Analyst Incidents

  • Possible HTTP Command and Control
  • Possible HTTP Command and Control to Multiple Endpoints
  • Unusual Repeated Connections
  • Unusual Repeated Connections to Multiple Endpoints
  • Multiple DNS Requests for Algorithmically Generated Domains

Indicators of Compromise

IoC - Type - Description

185.141.63[.]172 - IP Address - Socks5Systemz C2 Endpoint

193.242.211[.]141 - IP Address - Socks5Systemz C2 Endpoint

109.230.199[.]181 - IP Address - Socks5Systemz C2 Endpoint

109.236.88[.]134 - IP Address - Socks5Systemz C2 Endpoint

217.23.5[.]14 - IP Address - Socks5Systemz Proxy Checker App

88.80.148[.]8 - IP Address - Socks5Systemz Backconnect Endpoint

88.80.148[.]219 - IP Address - Socks5Systemz Backconnect Endpoint

185.141.63[.]4 - IP Address - Socks5Systemz Backconnect Endpoint

185.141.63[.]2 - IP Address - Socks5Systemz Backconnect Endpoint

195.154.188[.]211 - IP Address - Socks5Systemz Backconnect Endpoint

91.92.111[.]132 - IP Address - Socks5Systemz Backconnect Endpoint

91.121.30[.]185 - IP Address - Socks5Systemz Backconnect Endpoint

94.23.58[.]173 - IP Address - Socks5Systemz Backconnect Endpoint

37.187.148[.]204 - IP Address - Socks5Systemz Backconnect Endpoint

188.165.192[.]18 - IP Address - Socks5Systemz Backconnect Endpoint

/single.php?c=<RC4 data hex encoded> - URI - Socks5Systemz HTTP GET Request

/sign/<RC4 data hex encoded> - URI - Socks5Systemz HTTP GET Request

/proxy-activity.txt - URI - Socks5Systemz HTTP GET Request

datasheet[.]fun - Hostname - Socks5Systemz C2 Endpoint

bddns[.]cc - Hostname - Socks5Systemz C2 Endpoint

send-monitoring[.]bit - Hostname - Socks5Systemz C2 Endpoint

MITRE ATT&CK Mapping

Command and Control

T1071 - Application Layer Protocol

T1071.001 – Web protocols

T1568 – Dynamic Resolution

T1568.002 – Domain Generation Algorithms

T1132 – Data Encoding

T1132 – Non-Standard Encoding

T1090 – Proxy

T1090.002 – External Proxy

Exfiltration

T1041 – Exfiltration over C2 channel

Impact

T1496 – Resource Hijacking

References

1. https://www.bleepingcomputer.com/news/security/socks5systemz-proxy-service-infects-10-000-systems-worldwide/

2. https://www.bitsight.com/blog/unveiling-socks5systemz-rise-new-proxy-service-privateloader-and-amadey

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
Adam Potter
Senior Cyber 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