Blog
/
Network
/
September 4, 2022

Steps of a BumbleBee Intrusion to a Cobalt Strike

Discover the steps of a Bumblebee intrusion, from initial detection to Cobalt Strike deployment. Learn how Darktrace defends against evolving threats with AI.
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
Sam Lister
Specialist Security Researcher
Default blog image
04
Sep 2022

Introduction

Throughout April 2022, Darktrace observed several cases in which threat actors used the loader known as ‘BumbleBee’ to install Cobalt Strike Beacon onto victim systems. The threat actors then leveraged Cobalt Strike Beacon to conduct network reconnaissance, obtain account password data, and write malicious payloads across the network. In this article, we will provide details of the actions threat actors took during their intrusions, as well as details of the network-based behaviours which served as evidence of the actors’ activities.  

BumbleBee 

In March 2022, Google’s Threat Analysis Group (TAG) provided details of the activities of an Initial Access Broker (IAB) group dubbed ‘Exotic Lily’ [1]. Before March 2022, Google’s TAG observed Exotic Lily leveraging sophisticated impersonation techniques to trick employees of targeted organisations into downloading ISO disc image files from legitimate file storage services such as WeTransfer. These ISO files contained a Windows shortcut LNK file and a BazarLoader Dynamic Link Library (i.e, DLL). BazarLoader is a member of the Bazar family — a family of malware (including both BazarLoader and BazarBackdoor) with strong ties to the Trickbot malware, the Anchor malware family, and Conti ransomware. BazarLoader, which is typically distributed via email campaigns or via fraudulent call campaigns, has been known to drop Cobalt Strike as a precursor to Conti ransomware deployment [2]. 

In March 2022, Google’s TAG observed Exotic Lily leveraging file storage services to distribute an ISO file containing a DLL which, when executed, caused the victim machine to make HTTP requests with the user-agent string ‘bumblebee’. Google’s TAG consequently called this DLL payload ‘BumbleBee’. Since Google’s discovery of BumbleBee back in March, several threat research teams have reported BumbleBee samples dropping Cobalt Strike [1]/[3]/[4]/[5]. It has also been reported by Proofpoint [3] that other threat actors such as TA578 and TA579 transitioned to BumbleBee in March 2022.  

Interestingly, BazarLoader’s replacement with BumbleBee seems to coincide with the leaking of the Conti ransomware gang’s Jabber chat logs at the end of February 2022. On February 25th, 2022, the Conti gang published a blog post announcing their full support for the Russian state’s invasion of Ukraine [6]. 

Figure 1: The Conti gang's public declaration of their support for Russia's invasion of Ukraine

Within days of sharing their support for Russia, logs from a server hosting the group’s Jabber communications began to be leaked on Twitter by @ContiLeaks [7]. The leaked logs included records of conversations among nearly 500 threat actors between Jan 2020 and March 2022 [8]. The Jabber logs were supposedly stolen and leaked by a Ukrainian security researcher [3]/[6].

Affiliates of the Conti ransomware group were known to use BazarLoader to deliver Conti ransomware [9]. BumbleBee has now also been linked to the Conti ransomware group by several threat research teams [1]/[10]/[11]. The fact that threat actors’ transition from BazarLoader to BumbleBee coincides with the leak of Conti’s Jabber chat logs may indicate that the transition occurred as a result of the leaks [3]. Since the transition, BumbleBee has become a significant tool in the cyber-crime ecosystem, with links to several ransomware operations such as Conti, Quantum, and Mountlocker [11]. The rising use of BumbleBee by threat actors, and particularly ransomware actors, makes the early detection of BumbleBee key to identifying the preparatory stages of ransomware attacks.  

Intrusion Kill Chain 

In April 2022, Darktrace observed the following pattern of threat actor activity within the networks of several Darktrace clients: 

1.     Threat actor socially engineers user via email into running a BumbleBee payload on their device

2.     BumbleBee establishes HTTPS communication with a BumbleBee C2 server

3.     Threat actor instructs BumbleBee to download and execute Cobalt Strike Beacon

4.     Cobalt Strike Beacon establishes HTTPS communication with a Cobalt Strike C2 server

5.     Threat actor instructs Cobalt Strike Beacon to scan for open ports and to enumerate network shares

6.     Threat actor instructs Cobalt Strike Beacon to use the DCSync technique to obtain password account data from an internal domain controller

7.     Threat actor instructs Cobalt Strike Beacon to distribute malicious payloads to other internal systems 

With limited visibility over affected clients’ email environments, Darktrace was unable to determine how the threat actors interacted with users to initiate the BumbleBee infection. However, based on open-source reporting on BumbleBee [3]/[4]/[10]/[11]/[12]/[13]/[14]/[15]/[16]/[17], it is likely that the actors tricked target users into running BumbleBee by sending them emails containing either a malicious zipped ISO file or a link to a file storage service hosting the malicious zipped ISO file. These ISO files typically contain a LNK file and a BumbleBee DLL payload. The properties of these LNK files are set in such a way that opening them causes the corresponding DLL payload to run. 

In several cases observed by Darktrace, devices contacted a file storage service such as Microsoft OneDrive or Google Cloud Storage immediately before they displayed signs of BumbleBee infection. In these cases, it is likely that BumbleBee was executed on the users’ devices as a result of the users interacting with an ISO file which they were tricked into downloading from a file storage service. 

Figure 2: The above figure, taken from the event log for an infected device, shows that the device contacted a OneDrive endpoint immediately before making HTTPS connections to the BumbleBee C2 server, 45.140.146[.]244
Figure 3: The above figure, taken from the event log for an infected device, shows that the device contacted a Google Cloud Storage endpoint and then the malicious endpoint ‘marebust[.]com’ before making HTTPS connections to the  BumbleBee C2 servers, 108.62.118[.]61 and 23.227.198[.]217

After users ran a BumbleBee payload, their devices immediately initiated communications with BumbleBee C2 servers. The BumbleBee samples used HTTPS for their C2 communication, and all presented a common JA3 client fingerprint, ‘0c9457ab6f0d6a14fc8a3d1d149547fb’. All analysed samples excluded domain names in their ‘client hello’ messages to the C2 servers, which is unusual for legitimate HTTPS communication. External SSL connections which do not specify a destination domain name and whose JA3 client fingerprint is ‘0c9457ab6f0d6a14fc8a3d1d149547fb’ are potential indicators of BumbleBee infection. 

Figure 4:The above figure, taken from Darktrace's Advanced Search interface, depicts an infected device's spike in HTTPS connections with the JA3 client fingerprint ‘0c9457ab6f0d6a14fc8a3d1d149547fb’

Once the threat actors had established HTTPS communication with the BumbleBee-infected systems, they instructed BumbleBee to download and execute Cobalt Strike Beacon. This behaviour resulted in the infected systems making HTTPS connections to Cobalt Strike C2 servers. The Cobalt Strike Beacon samples all had the same JA3 client fingerprint ‘a0e9f5d64349fb13191bc781f81f42e1’ — a fingerprint associated with previously seen Cobalt Strike samples [18]. The domain names ‘fuvataren[.]com’ and ‘cuhirito[.]com’ were observed in the samples’ HTTPS communications. 

Figure 5:The above figure, taken from Darktrace's Advanced Search interface, depicts the Cobalt Strike C2 communications which immediately followed a device's BumbleBee C2 activity

Cobalt Strike Beacon payloads call home to C2 servers for instructions. In the cases observed, threat actors first instructed the Beacon payloads to perform reconnaissance tasks, such as SMB port scanning and SMB enumeration. It is likely that the threat actors performed these steps to inform the next stages of their operations.  The SMB enumeration activity was evidenced by the infected devices making NetrShareEnum and NetrShareGetInfo requests to the srvsvc RPC interface on internal systems.

Figure 6: The above figure, taken from Darktrace’s Advanced Search interface, depicts a spike in srvsvc requests coinciding with the infected device's Cobalt Strike C2 activity

After providing Cobalt Strike Beacon with reconnaissance tasks, the threat actors set out to obtain account password data in preparation for the lateral movement phase of their operation. To obtain account password data, the actors instructed Cobalt Strike Beacon to use the DCSync technique to replicate account password data from an internal domain controller. This activity was evidenced by the infected devices making DRSGetNCChanges requests to the drsuapi RPC interface on internal domain controllers. 

Figure 7: The above figure, taken from Darktrace’s Advanced Search interface, depicts a spike in DRSGetNCChanges requests coinciding with the infected device’s Cobalt Strike C2 activity

After leveraging the DCSync technique, the threat actors sought to broaden their presence within the targeted networks.  To achieve this, they instructed Cobalt Strike Beacon to get several specially selected internal systems to run a suspiciously named DLL (‘f.dll’). Cobalt Strike first established SMB sessions with target systems using compromised account credentials. During these sessions, Cobalt Strike uploaded the malicious DLL to a hidden network share. To execute the DLL, Cobalt Strike abused the Windows Service Control Manager (SCM) to remotely control and manipulate running services on the targeted internal hosts. Cobalt Strike first opened a binding handle to the svcctl interface on the targeted destination systems. It then went on to make an OpenSCManagerW request, a CreateServiceA request, and a StartServiceA request to the svcctl interface on the targeted hosts: 

·      Bind request – opens a binding handle to the relevant RPC interface (in this case, the svcctl interface) on the destination device

·      OpenSCManagerW request – establishes a connection to the Service Control Manager (SCM) on the destination device and opens a specified SCM database

·      CreateServiceA request – creates a service object and adds it to the specified SCM database 

·      StartServiceA request – starts a specified service

Figure 8: The above figure, taken from Darktrace’s Advanced Search interface, outlines an infected system’s lateral movement activities. After writing a file named ‘f.dll’ to the C$ share on an internal server, the infected device made several RPC requests to the svcctl interface on the targeted server

It is likely that the DLL file which the threat actors distributed was a Cobalt Strike payload. In one case, however, the threat actor was also seen distributing and executing a payload named ‘procdump64.exe’. This may suggest that the threat actor was seeking to use ProcDump to obtain authentication material stored in the process memory of the Local Security Authority Subsystem Service (LSASS). Given that ProcDump is a legitimate Windows Sysinternals tool primarily used for diagnostics and troubleshooting, it is likely that threat actors leveraged it in order to evade detection. 

In all the cases which Darktrace observed, threat actors’ attempts to conduct follow-up activities after moving laterally were thwarted with the help of Darktrace’s SOC team. It is likely that the threat actors responsible for the reported activities were seeking to deploy ransomware within the targeted networks. The steps which the threat actors took to make progress towards achieving this objective resulted in highly unusual patterns of network traffic. Darktrace’s detection of these unusual network activities allowed security teams to prevent these threat actors from achieving their disruptive objectives. 

Darktrace Coverage

Once threat actors succeeded in tricking users into running BumbleBee on their devices, Darktrace’s Self-Learning AI immediately detected the command-and-control (C2) activity generated by the loader. BumbleBee’s C2 activity caused the following Darktrace models to breach:

·      Anomalous Connection / Anomalous SSL without SNI to New External

·      Anomalous Connection / Suspicious Self-Signed SSL

·      Anomalous Connection / Rare External SSL Self-Signed

·      Compromise / Suspicious TLS Beaconing To Rare External

·      Compromise / Beacon to Young Endpoint

·      Compromise / Beaconing Activity To External Rare

·      Compromise / Sustained SSL or HTTP Increase

·      Compromise / Suspicious TLS Beaconing To Rare External

·      Compromise / SSL Beaconing to Rare Destination

·      Compromise / Large Number of Suspicious Successful Connections

·      Device / Multiple C2 Model Breaches 

BumbleBee’s delivery of Cobalt Strike Beacon onto target systems resulted in those systems communicating with Cobalt Strike C2 servers. Cobalt Strike Beacon’s C2 communications resulted in breaches of the following models: 

·      Compromise / Beaconing Activity To External Rare

·      Compromise / High Volume of Connections with Beacon Score

·      Compromise / Large Number of Suspicious Successful Connections

·      Compromise / Sustained SSL or HTTP Increase

·      Compromise / SSL or HTTP Beacon

·      Compromise / Slow Beaconing Activity To External Rare

·      Compromise / SSL Beaconing to Rare Destination 

The threat actors’ subsequent port scanning and SMB enumeration activities caused the following models to breach:

·      Device / Network Scan

·      Anomalous Connection / SMB Enumeration

·      Device / Possible SMB/NTLM Reconnaissance

·      Device / Suspicious Network Scan Activity  

The threat actors’ attempts to obtain account password data from domain controllers using the DCSync technique resulted in breaches of the following models: 

·      Compromise / Unusual SMB Session and DRS

·      Anomalous Connection / Anomalous DRSGetNCChanges Operation

Finally, the threat actors’ attempts to internally distribute and execute payloads resulted in breaches of the following models:

·      Compliance / SMB Drive Write

·      Device / Lateral Movement and C2 Activity

·      Device / SMB Lateral Movement

·      Device / Multiple Lateral Movement Model Breaches

·      Anomalous File / Internal / Unusual SMB Script Write

·      Anomalous File / Internal / Unusual Internal EXE File Transfer

·      Anomalous Connection / High Volume of New or Uncommon Service Control

If Darktrace/Network had been configured in the targeted environments, then it would have blocked BumbleBee’s C2 communications, which would have likely prevented the threat actors from delivering Cobalt Strike Beacon into the target networks. 

Figure 9: Attack timeline

Conclusion

Threat actors use loaders to smuggle more harmful payloads into target networks. Prior to March 2022, it was common to see threat actors using the BazarLoader loader to transfer their payloads into target environments. However, since the public disclosure of the Conti gang’s Jabber chat logs at the end of February, the cybersecurity world has witnessed a shift in tradecraft. Threat actors have seemingly transitioned from using BazarLoader to using a novel loader known as ‘BumbleBee’. Since BumbleBee first made an appearance in March 2022, a growing number of threat actors, in particular ransomware actors, have been observed using it.

It is likely that this trend will continue, which makes the detection of BumbleBee activity vital for the prevention of ransomware deployment within organisations’ networks. During April, Darktrace’s SOC team observed a particular pattern of threat actor activity involving the BumbleBee loader. After tricking users into running BumbleBee on their devices, threat actors were seen instructing BumbleBee to drop Cobalt Strike Beacon. Threat actors then leveraged Cobalt Strike Beacon to conduct network reconnaissance, obtain account password data from internal domain controllers, and distribute malicious payloads internally.  Darktrace’s detection of these activities prevented the threat actors from achieving their likely harmful objectives.  

Thanks to Ross Ellis for his contributions to this blog.‍

Appendices 

References 

[1] https://blog.google/threat-analysis-group/exposing-initial-access-broker-ties-conti/ 

[2] https://securityintelligence.com/posts/trickbot-gang-doubles-down-enterprise-infection/ 

[3] https://www.proofpoint.com/us/blog/threat-insight/bumblebee-is-still-transforming

[4] https://www.cynet.com/orion-threat-alert-flight-of-the-bumblebee/ 

[5] https://research.nccgroup.com/2022/04/29/adventures-in-the-land-of-bumblebee-a-new-malicious-loader/ 

[6] https://www.bleepingcomputer.com/news/security/conti-ransomwares-internal-chats-leaked-after-siding-with-russia/ 

[7] https://therecord.media/conti-leaks-the-panama-papers-of-ransomware/ 

[8] https://www.secureworks.com/blog/gold-ulrick-leaks-reveal-organizational-structure-and-relationships 

[9] https://www.prodaft.com/m/reports/Conti_TLPWHITE_v1.6_WVcSEtc.pdf 

[10] https://www.kroll.com/en/insights/publications/cyber/bumblebee-loader-linked-conti-used-in-quantum-locker-attacks 

[11] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/bumblebee-loader-cybercrime 

[12] https://isc.sans.edu/diary/TA578+using+thread-hijacked+emails+to+push+ISO+files+for+Bumblebee+malware/28636 

[13] https://isc.sans.edu/diary/rss/28664 

[14] https://www.logpoint.com/wp-content/uploads/2022/05/buzz-of-the-bumblebee-a-new-malicious-loader-threat-report-no-3.pdf 

[15] https://ghoulsec.medium.com/mal-series-23-malware-loader-bumblebee-6ab3cf69d601 

[16]  https://blog.cyble.com/2022/06/07/bumblebee-loader-on-the-rise/  

[17]  https://asec.ahnlab.com/en/35460/ 

[18] https://thedfirreport.com/2021/07/19/icedid-and-cobalt-strike-vs-antivirus/

‍

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
Sam Lister
Specialist Security Researcher

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