Blog
/
Network
/
July 4, 2024

A Busy Agenda: Darktrace's Detection of Qilin Ransomware as a Service Operator

This blog breaks down how Darktrace detected and analyzed Qilin, a Ransomware-as-a-Service group behind recent high-impact attacks. You’ll see how Qilin affiliates customize attacks with flexible encryption, process termination, and double-extortion techniques, as well as why its cross-platform builds in Rust and Golang make it especially evasive. Darktrace highlights three real-world cases where its AI identified likely Qilin activity across customer environments, offering insights into how behavioral detection can spot novel ransomware before disruption occurs. Readers will gain a clear view of Qilin’s toolkit, tactics, and how self-learning defense adapts to these evolving threats.
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
Alexandra Sentenac
Cyber Analyst
Default blog image
04
Jul 2024

Qilin ransomware has recently dominated discussions across the cyber security landscape following its deployment in an attack on Synnovis, a UK-based medical laboratory company. The ransomware attack ultimately affected patient services at multiple National Health Service (NHS) hospitals that rely on Synnovis diagnostic and pathology services. Qilin’s origins, however, date back further to October 2022 when the group was observed seemingly posting leaked data from its first known victim on its Dedicated Leak Site (DLS) under the name Agenda[1].

The Darktrace Threat Research team investigated network artifacts related to Qilin and identified three probable cases of the ransomware across the Darktrace customer base between June 2022 and May 2024.

Qilin Ransomware-as-a-Service Operator

Qilin operates as a Ransomware-as-a-Service (RaaS) that employs double extortion tactics, whereby harvested data is exfiltrated and threatened of publication on the group's DLS, which is hosted on Tor. Qilin ransomware has samples written in both the Golang and Rust programming languages, making it compilable with various operating systems, and is highly customizable. When building Qilin ransomware variants to be used on their target(s), affiliates can configure settings such as the encryption mode (i.e., skip-step, percent, and speed), the file extension being appended, files, extensions and directories to be skipped during the encryption, and the processes and services to be terminated, among others[1] [2].  

Trend Micro analysts, who were the first to discover Qilin samples in August 2022, when the name "Agenda" was still used in ransom notes, found that each analyzed sample was customized for the intended victims and that "unique company IDs were used as extensions of encrypted files" [3]. This information is configurable from within the Qilin's affiliate panel's 'Targets' section, shown below. The panel's background image features the eponym Chinese legendary chimerical creature Qilin (pronounced “Ke Lin”). Despite this Chinese mythology reference, Russian language was observed being used by a Qilin operator in an underground forum post aimed at hiring affiliates and advertising their RaaS operation[2].

Figure 1: Qilin ransomware’s affiliate panel.

Qilin's RaaS program purportedly has an attractive affiliates' payment structure, with affiliates allegedly able to earn 80% of ransom payments of USD 3m or less and 85% for payments above that figure[2], making it a possibly appealing option in the RaaS ecosystem.  Publication of stolen data and ransom payment negotiations are purportedly handled by Qilin operators. Qilin affiliates have been known to target companies located around the world and within a variety of industries, including critical sectors such as healthcare and energy.

As Qilin is a RaaS operation, the choice of targets does not necessarily reflect Qilin operators' intentions, but rather that of its affiliates.  Similarly, the tactics, techniques, procedures (TTPs) and indicators of compromise (IoC) identified by Darktrace are associated with the given affiliate deploying Qilin ransomware for their own purpose, rather than TTPs and IoCs of the Qilin group. Likewise, initial vectors of infection may vary from affiliate to affiliate. Previous studies show that initial access to networks were gained via spear phishing emails or by leveraging exposed applications and interfaces.

Differences have been observed in terms of data exfiltration and potential C2 external endpoints, suggesting the below investigations are not all related to the same group or actor(s).

Darktrace’s Threat Research Investigation

June 2022

Darktrace first detected an instance of Qilin ransomware back in June 2022, when an attacker was observed successfully accessing a customer’s Virtual Private Network (VPN) and compromising an administrative account, before using RDP to gain access to the customer’s Microsoft System Center Configuration Manager (SCCM) server

From there, an attack against the customer's VMware ESXi hosts was launched. Fortunately, a reboot of their virtual machines (VM) caught the attention of the security team who further uncovered that custom profiles had been created and remote scripts executed to change root passwords on their VM hosts. Three accounts were found to have been compromised and three systems encrypted by ransomware.  

Unfortunately, Darktrace was not configured to monitor the affected subnets at the time of the attack. Despite this, the customer was able to work directly with Darktrace analysts via the Ask the Expert (ATE) service to add the subnets in question to Darktrace’s visibility, allowing it to monitor for any further unusual behavior.

Once visibility over the compromised SCCM server was established, Darktrace observed a series of unusual network scanning activities and the use of Kali (a Linux distribution designed for digital forensics and penetration testing). Furthermore, the server was observed making connections to multiple rare external hosts, many using the “[.]ru” Top Level Domain (TLD). One of the external destinations the server was attempting to connect was found to be related to SystemBC, a malware that turns infected hosts into SOCKS5 proxy bots and provides command-and-control (C2) functionality.

Additionally, the server was observed making external connections over ports 993 and 143 (typically associated with the use of the Interactive Message Access Protocol (IMAP) to multiple rare external endpoints. This was likely due to the presence of Tofsee malware on the device.

After the compromise had been contained, Darktrace identified several ransom notes following the naming convention “README-RECOVER-<extension/company_id>.txt”” on the network. This naming convention, as well as the similar “<company_id>-RECOVER-README.txt” have been referenced by open-source intelligence (OSINT) providers as associated with Qilin ransom notes[5] [6] [7].

April 2023

The next case of Qilin ransomware observed by Darktrace took place in April 2023 on the network of a customer in the manufacturing sector in APAC. Unfortunately for the customer in this instance, Darktrace RESPOND™ was not active on their environment and no autonomous response actions were taken to contain the compromise.

Over the course of two days, Darktrace identified a wide range of malicious activity ranging from extensive initial scanning and lateral movement attempts to the writing of ransom notes that followed the aforementioned naming convention (i.e., “README-RECOVER-<extension/company_id>.txt”).

Darktrace observed two affected devices attempting to move laterally through the SMB, DCE-RPC and RDP network protocols. Default credentials (e.g., UserName, admin, administrator) were also observed in the large volumes of SMB sessions initiated by these devices. One of the target devices of these SMB connections was a domain controller, which was subsequently seen making suspicious WMI requests to multiple devices over DCE-RPC and enumerating SMB shares by binding to the ‘server service’ (srvsvc) named pipe to a high number of internal devices within a short time frame. The domain controller was further detected establishing an anomalously high number of connections to several internal devices, notably using the RDP administrative protocol via a default admin cookie.  

Repeated connections over the HTTP and SSL protocol to multiple newly observed IPs located in the 184.168.123.0/24 range were observed, indicating C2 connectivity.  WebDAV user agent and a JA3 fingerprint potentially associated with Cobalt Strike were notably observed in these connections. A few hours later, Darktrace detected additional suspicious external connections, this time to IPs associated with the MEGA cloud storage solution. Storage solutions such as MEGA are often abused by attackers to host stolen data post exfiltration. In this case, the endpoints were all rare for the network, suggesting this solution was not commonly used by legitimate users. Around 30 GB of data was exfiltrated over the SSL protocol.

Darktrace did not observe any encryption-related activity on this customer’s network, suggesting that encryption may have taken place locally or within network segments not monitored by Darktrace.

May 2024

The most recent instance of Qilin observed by Darktrace took place in May 2024 and involved a customer in the US. In this case, Darktrace initially detected affected devices using unusual administrative and default credentials, before additional internal systems were observed making extensive suspicious DCE-RPC requests to a range of internal locations, performing network scanning, making unusual internal RDP connections, and transferring suspicious executable files like 'a157496.exe' and '83b87b2.exe'.  SMB writes of the file "LSM_API_service" were also observed, activity which was considered 100% unusual by Darktrace; this is an RPC service that can be abused to enumerate logged-in users and steal their tokens. Various repeated connections likely representative of C2 communications were detected via both HTTP and SSL to rare external endpoints linked in OSINT to Cobalt Strike use. During these connections, HTTP GET requests for the following URIs were observed:

/asdffHTTPS

/asdfgdf

/asdfgHTTP

/download/sihost64.dll

Notably, this included a GET request a DLL file named "sihost64.dll" from a domain controller using PowerShell.  

Over 102 GB of data may have been transferred to another previously unseen endpoint, 194.165.16[.]13, via the unencrypted File Transfer Protocol (FTP). Additionally, many non-FTP connections to the endpoint could be observed, over which more than 783 GB of data was exfiltrated. Regarding file encryption activity, a wide range of destination devices and shares were targeted.

Figure 2: Advanced Search graph displaying the total volume of data transferred over FTP to a malicious IP.

During investigations, Darktrace’s Threat Research team identified an additional customer, also based in the United States, where similar data exfiltration activity was observed in April 2024. Although no indications of ransomware encryption were detected on the network, multiple similarities were observed with the case discussed just prior. Notably, the same exfiltration IP and protocol (194.165.16[.]13 and FTP, respectively) were identified in both cases. Additional HTTP connectivity was further observed to another IP using a self-signed certificate (i.e., CN=ne[.]com,OU=key operations,O=1000,L=,ST=,C=KM) located within the same ASN (i.e., AS48721 Flyservers S.A.). Some of the URIs seen in the GET requests made to this endpoint were the same as identified in that same previous case.

Information regarding another device also making repeated connections to the same IP was described in the second event of the same Cyber AI Analyst incident. Following this C2 connectivity, network scanning was observed from a compromised domain controller, followed by additional reconnaissance and lateral movement over the DCE-RPC and SMB protocols. Darktrace again observed SMB writes of the file "LSM_API_service", as in the previous case, activity which was also considered 100% unusual for the network. These similarities suggest the same actor or affiliate may have been responsible for activity observed, even though no encryption was observed in the latter case.

Figure 3: First event of the Cyber AI Analyst investigation following the compromise activity.

According to researchers at Microsoft, some of the IoCs observed on both affected accounts are associated with Pistachio Tempest, a threat actor reportedly associated with ransomware distribution. The Microsoft threat actor naming convention uses the term "tempest" to reference criminal organizations with motivations of financial gain that are not associated with high confidence to a known non-nation state or commercial entity. While Pistachio Tempest’s TTPs have changed over time, their key elements still involve ransomware, exfiltration, and extortion. Once they've gained access to an environment, Pistachio Tempest typically utilizes additional tools to complement their use of Cobalt Strike; this includes the use of the SystemBC RAT and the SliverC2 framework, respectively. It has also been reported that Pistacho Tempest has experimented with various RaaS offerings, which recently included Qilin ransomware[4].

Conclusion

Qilin is a RaaS group that has gained notoriety recently due to high-profile attacks perpetrated by its affiliates. Despite this, the group likely includes affiliates and actors who were previously associated with other ransomware groups. These individuals bring their own modus operandi and utilize both known and novel TTPs and IoCs that differ from one attack to another.

Darktrace’s anomaly-based technology is inherently threat-agnostic, treating all RaaS variants equally regardless of the attackers’ tools and infrastructure. Deviations from a device’s ‘learned’ pattern of behavior during an attack enable Darktrace to detect and contain potentially disruptive ransomware attacks.

Credit to: Alexandra Sentenac, Emma Foulger, Justin Torres, Min Kim, Signe Zaharka for their contributions.

References

[1] https://www.sentinelone.com/anthology/agenda-qilin/  

[2] https://www.group-ib.com/blog/qilin-ransomware/

[3] https://www.trendmicro.com/en_us/research/22/h/new-golang-ransomware-agenda-customizes-attacks.html

[4] https://www.microsoft.com/en-us/security/security-insider/pistachio-tempest

[5] https://www.trendmicro.com/en_us/research/22/h/new-golang-ransomware-agenda-customizes-attacks.html

[6] https://www.bleepingcomputer.com/forums/t/790240/agenda-qilin-ransomware-id-random-10-char;-recover-readmetxt-support/

[7] https://github.com/threatlabz/ransomware_notes/tree/main/qilin

Darktrace Model Detections

Internal Reconnaissance

Device / Suspicious SMB Scanning Activity

Device / Network Scan

Device / RDP Scan

Device / ICMP Address Scan

Device / Suspicious Network Scan Activity

Anomalous Connection / SMB Enumeration

Device / New or Uncommon WMI Activity

Device / Attack and Recon Tools

Lateral Movement

Device / SMB Session Brute Force (Admin)

Device / Large Number of Model Breaches from Critical Network Device

Device / Multiple Lateral Movement Model Breaches

Anomalous Connection / Unusual Admin RDP Session

Device / SMB Lateral Movement

Compliance / SMB Drive Write

Anomalous Connection / New or Uncommon Service Control

Anomalous Connection / Anomalous DRSGetNCChanges Operation

Anomalous Server Activity / Domain Controller Initiated to Client

User / New Admin Credentials on Client

C2 Communication

Anomalous Server Activity / Outgoing from Server

Anomalous Connection / Multiple Connections to New External TCP Port

Anomalous Connection / Anomalous SSL without SNI to New External

Anomalous Connection / Rare External SSL Self-Signed

Device / Increased External Connectivity

Unusual Activity / Unusual External Activity

Compromise / New or Repeated to Unusual SSL Port

Anomalous Connection / Multiple Failed Connections to Rare Endpoint

Device / Suspicious Domain

Device / Increased External Connectivity

Compromise / Sustained SSL or HTTP Increase

Compromise / Botnet C2 Behaviour

Anomalous Connection / POST to PHP on New External Host

Anomalous Connection / Multiple HTTP POSTs to Rare Hostname

Anomalous File / EXE from Rare External Location

Exfiltration

Unusual Activity / Enhanced Unusual External Data Transfer

Anomalous Connection / Data Sent to Rare Domain

Unusual Activity / Unusual External Data Transfer

Anomalous Connection / Uncommon 1 GiB Outbound

Unusual Activity / Unusual External Data to New Endpoint

Compliance / FTP / Unusual Outbound FTP

File Encryption

Compromise / Ransomware / Suspicious SMB Activity

Anomalous Connection / Sustained MIME Type Conversion

Anomalous File / Internal / Additional Extension Appended to SMB File

Compromise / Ransomware / Possible Ransom Note Write

Compromise / Ransomware / Possible Ransom Note Read

Anomalous Connection / Suspicious Read Write Ratio

IoC List

IoC – Type – Description + Confidence

93.115.25[.]139 IP C2 Server, likely associated with SystemBC

194.165.16[.]13 IP Probable Exfiltration Server

91.238.181[.]230 IP C2 Server, likely associated with Cobalt Strike

ikea0[.]com Hostname C2 Server, likely associated with Cobalt Strike

lebondogicoin[.]com Hostname C2 Server, likely associated with Cobalt Strike

184.168.123[.]220 IP Possible C2 Infrastructure

184.168.123[.]219 IP Possible C2 Infrastructure

184.168.123[.]236 IP Possible C2 Infrastructure

184.168.123[.]241 IP Possible C2 Infrastructure

184.168.123[.]247 IP Possible C2 Infrastructure

184.168.123[.]251 IP Possible C2 Infrastructure

184.168.123[.]252 IP Possible C2 Infrastructure

184.168.123[.]229 IP Possible C2 Infrastructure

184.168.123[.]246 IP Possible C2 Infrastructure

184.168.123[.]230 IP Possible C2 Infrastructure

gfs440n010.userstorage.me ga.co[.]nz Hostname Possible Exfiltration Server. Not inherently malicious; associated with MEGA file storage.

gfs440n010.userstorage.me ga.co[.]nz Hostname Possible Exfiltration Server. Not inherently malicious; associated with MEGA file storage.

Get the latest insights on emerging cyber threats

This report explores the latest trends shaping the cybersecurity landscape and what defenders need to know in 2025

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
Alexandra Sentenac
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