ブログ
/
Network
/
October 26, 2022

Strategies to Prolong Quantum Ransomware Attacks

Learn more about how Darktrace combats Quantum Ransomware changing strategy for cyberattacks. Explore the power of AI-driven network cyber security!
Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Nicole Wong
Cyber Security Analyst
Default blog image
26
Oct 2022

Within science and engineering, the word ‘quantum’ may spark associations with speed and capability, referencing a superior computer that can perform tasks a classical computer cannot. In cyber security, some may recognize ‘quantum’ in relation to cryptography or, more recently, as the name of a new ransomware group, which achieved network-wide encryption a mere four hours after an initial infection.   

Although this group now has a reputation for carrying out fast and efficient attacks, speed is not their only tactic. In August 2022, Darktrace detected a Quantum Ransomware incident where attackers remained in the victim’s network for almost a month after the initial signs of infection, before detonating ransomware. This was a stark difference to previously reported attacks, demonstrating that as motives change, so do threat actors’ strategies. 

The Quantum Group

Quantum was first identified in August 2021 as the latest of several rebrands of MountLocker ransomware [1]. As part of this rebrand, the extension ‘.quantum’ is appended to filenames that are encrypted and the associated ransom notes are named ‘README_TO_DECRYPT.html’ [2].  

From April 2022, media coverage of this group has increased following a DFIR report detailing an attack that progressed from initial access to domain-wide ransomware within four hours [3]. To put this into perspective, the global median dwell time for ransomware in 2020 and 2021 is 5 days [4]. In the case of Quantum, threat actors gained direct keyboard access to devices merely 2 hours after initial infection. The ransomware was staged on the domain controller around an hour and a half later, and executed 12 minutes after that.   

Quantum’s behaviour bears similarities to other groups, possibly due to their history and recruitment. Several members of the disbanded Conti ransomware group are reported to have joined the Quantum and BumbleBee operations. Security researchers have also identified similarities in the payloads and C2 infrastructure used by these groups [5 & 6].  Notably, these are the IcedID initial payload and Cobalt Strike C2 beacon used in this attack. Darktrace has also observed and prevented IcedID and Cobalt Strike activity from BumbleBee across several customer environments.

The Attack

From 11th July 2022, a device suspected to be patient zero made repeated DNS queries for external hosts that appear to be associated with IcedID C2 traffic [7 & 8]. In several reported cases [9 & 10], this banking trojan is delivered through a phishing email containing a malicious attachment that loads an IcedID DLL. As Darktrace was not deployed in the prospect’s email environment, there was no visibility of the initial access vector, however an example of a phishing campaign containing this payload is presented below. It is also possible that the device was already infected prior to joining the network. 

Figure 1- An example phishing email used to distribute IcedID. If configured, Darktrace/Email would be able to detect that the email was sent from an anomalous sender, was part of a fake reply chain, and had a suspicious attachment containing compressed content of unusual mime type [11].    

 

Figure 2- The DNS queries to endpoints associated with IcedID C2 servers, taken from the infected device’s event log.  Additional DNS queries made to other IcedID C2 servers are in the list of IOCs in the appendices.  The repeated DNS queries are indicative of beaconing.

‍
It was not until 22nd July that activity was seen which indicated the attack had progressed to the next stage of the kill chain. This contrasts the previously seen attacks where the progression to Cobalt Strike C2 beaconing and reconnaissance and lateral movement occurred within 2 hours of the initial infection [12 & 13]. In this case, patient zero initiated numerous unusual connections to other internal devices using a compromised account, connections that were indicative of reconnaissance using built-in Windows utilities:

·      DNS queries for hostnames in the network

·      SMB writes to IPC$ shares of those hostnames queried, binding to the srvsvc named pipe to enumerate things such as SMB shares and services on a device, client access permissions on network shares and users logged in to a remote session

·      DCE-RPC connections to the endpoint mapper service, which enables identification of the ports assigned to a particular RPC service

These connections were initiated using an existing credential on the device and just like the dwelling time, differed from previously reported Quantum group attacks where discovery actions were spawned and performed automatically by the IcedID process [14]. Figure 3 depicts how Darktrace detected that this activity deviated from the device’s normal behaviour.  

Figure 3- This figure displays the spike in active internal connections initiated by patient zero. The coloured dots represent the Darktrace models that were breached, detecting this unusual reconnaissance and lateral movement activity.

‍

Four days later, on the 26th of July, patient zero performed SMB writes of DLL and MSI executables to the C$ shares of internal devices including domain controllers, using a privileged credential not previously seen on the patient zero device. The deviation from normal behaviour that this represents is also displayed in Figure 3. Throughout this activity, patient zero made DNS queries for the external Cobalt Strike C2 server shown in Figure 4. Cobalt Strike has often been seen as a secondary payload delivered via IcedID, due to IcedID’s ability to evade detection and deploy large scale campaigns [15]. It is likely that reconnaissance and lateral movement was performed under instructions received by the Cobalt Strike C2 server.   

Figure 4- This figure is taken from Darktrace’s Advanced Search interface, showing a DNS query for a Cobalt Strike C2 server occurring during SMB writes of .dll files and DCE-RPC requests to the epmapper service, demonstrating reconnaissance and lateral movement.

‍
The SMB writes to domain controllers and usage of a new account suggests that by this stage, the attacker had achieved domain dominance. The attacker also appeared to have had hands-on access to the network via a console; the repetition of the paths ‘programdata\v1.dll’ and ‘ProgramData\v1.dll’, in lower and title case respectively, suggests they were entered manually.  

These DLL files likely contained a copy of the malware that injects into legitimate processes such as winlogon, to perform commands that call out to C2 servers [16]. Shortly after the file transfers, the affected domain controllers were also seen beaconing to external endpoints (‘sezijiru[.]com’ and ‘gedabuyisi[.]com’) that OSINT tools have associated with these DLL files [17 & 18]. Moreover, these SSL connections were made using a default client fingerprint for Cobalt Strike [19], which is consistent with the initial delivery method. To illustrate the beaconing nature of these connections, Figure 5 displays the 4.3 million daily SSL connections to one of the C2 servers during the attack. The 100,000 most recent connections were initiated by 11 unique source IP addresses alone.

Figure 5- The Advanced Search interface, querying for external SSL connections from devices in the network to an external host that appears to be a Cobalt Strike C2 server. 4.3 million connections were made over 8 days, even after the ransomware was eventually detonated on 2022-08-03.


Shortly after the writes, the attack progressed to the penultimate stage. The next day, on the 27th of July, the attackers moved to achieve their first objective: data exfiltration. Data exfiltration is not always performed by the Quantum ransomware gang. Researchers have noted discrepancies between claims of data theft made in their ransom notes versus the lack of data seen leaving the network, although this may have been missed due to covert exfiltration via a Cobalt Strike beacon [20]. 

In contrast, this attack displayed several gigabytes of data leaving internal devices including servers that had previously beaconed to Cobalt Strike C2 servers. This data was transferred overtly via FTP, however the attacker still attempted to conceal the activity using ephemeral ports (FTP in EPSV mode). FTP is an effective method for attackers to exfiltrate large files as it is easy to use, organizations often neglect to monitor outbound usage, and it can be shipped through ports that will not be blocked by traditional firewalls [21].   

Figure 6 displays an example of the FTP data transfer to attacker-controlled infrastructure, in which the destination share appears structured to identify the organization that the data was stolen from, suggesting there may be other victim organizations’ data stored. This suggests that data exfiltration was an intended outcome of this attack. 

Figure 6- This figure is from Darktrace’s Advanced Search interface, displaying some of the data transferred from an internal device to the attacker’s FTP server.

 
Data was continuously exfiltrated until a week later when the final stage of the attack was achieved and Quantum ransomware was detonated. Darktrace detected the following unusual SMB activity initiated from the attacker-created account that is a hallmark for ransomware (see Figure 7 for example log):

·      Symmetric SMB Read to Write ratio, indicative of active encryption

·      Sustained MIME type conversion of files, with the extension ‘.quantum’ appended to filenames

·      SMB writes of a ransom note ‘README_TO_DECRYPT.html’ (see Figure 8 for an example note)

Figure 7- The Model Breach Event Log for a device that had files encrypted by Quantum ransomware, showing the reads and writes of files with ‘.quantum’ appended to encrypted files, and an HTML ransom note left where the files were encrypted.

 

Figure 8- An example of the ransom note left by the Quantum gang, this one is taken from open-sources [22].


The example in Figure 8 mentions that the attacker also possessed large volumes of victim data.  It is likely that the gigabytes of data exfiltrated over FTP were leveraged as blackmail to further extort the victim organization for payment.  

Darktrace Coverage

 

Figure 9- Timeline of Quantum ransomware incident

‍

If Darktrace/Email was deployed in the prospect’s environment, the initial payload (if delivered through a phishing email) could have been detected and held from the recipient’s inbox. Although DETECT identified anomalous network behaviour at each stage of the attack, since the incident occurred during a trial phase where Darktrace could only detect but not respond, the attack was able to progress through the kill chain. If RESPOND/Network had been configured in the targeted environment, the unusual connections observed during the initial access, C2, reconnaissance and lateral movement stages of the attack could have been blocked. This would have prevented the attackers from delivering the later stage payloads and eventual ransomware into the target network.

It is often thought that a properly implemented backup strategy is sufficient defense against ransomware [23], however as discussed in a previous Darktrace blog, the increasing frequency of double extortion attacks in a world where ‘data is the new oil’ demonstrates that backups alone are not a mitigation for the risk of a ransomware attack [24]. Equally, the lack of preventive defenses in the target’s environment enabled the attacker’s riskier decision to dwell in the network for longer and allowed them to optimize their potential reward. 

Recent crackdowns from law enforcement on ransomware groups have shifted these groups’ approaches to aim for a balance between low risk and significant financial rewards [25]. However, given the Quantum gang only have a 5% market share in Q2 2022, compared to the 13.2% held by LockBit and 16.9% held by BlackCat [26], a riskier strategy may be favourable, as a longer dwell time and double extortion outcome offers a ‘belt and braces’ approach to maximizing the rewards from carrying out this attack. Alternatively, the gaps in-between the attack stages may imply that more than one player was involved in this attack, although this group has not been reported to operate a franchise model before [27]. Whether assisted by others or driving for a risk approach, it is clear that Quantum (like other actors) are continuing to adapt to ensure their financial success. They will continue to be successful until organizations dedicate themselves to ensuring that the proper data protection and network security measures are in place. 

Conclusion 

Ransomware has evolved over time and groups have merged and rebranded. However, this incident of Quantum ransomware demonstrates that regardless of the capability to execute a full attack within hours, prolonging an attack to optimize potential reward by leveraging double extortion tactics is sometimes still the preferred action. The pattern of network activity mirrors the techniques used in other Quantum attacks, however this incident lacked the continuous progression of the group’s attacks reported recently and may represent a change of motives during the process. Knowing that attacker motives can change reinforces the need for organizations to invest in preventative controls- an organization may already be too far down the line if it is executing its backup contingency plans. Darktrace DETECT/Network had visibility over both the early network-based indicators of compromise and the escalation to the later stages of this attack. Had Darktrace also been allowed to respond, this case of Quantum ransomware would also have had a very short dwell time, but a far better outcome for the victim.

Thanks to Steve Robinson for his contributions to this blog.

Appendices

References

[1] https://community.ibm.com/community/user/security/blogs/tristan-reed/2022/07/13/ibm-security-reaqta-vs-quantum-locker-ransomware

 

[2] https://www.bleepingcomputer.com/news/security/quantum-ransomware-seen-deployed-in-rapid-network-attacks/

 

[3], [12], [14], [16], [20] https://thedfirreport.com/2022/04/25/quantum-ransomware/

 

[4] https://www.mandiant.com/sites/default/files/2022-04/M-Trends%202022%20Executive%20Summary.pdf

 

[5] https://cyware.com/news/over-650-healthcare-organizations-affected-by-the-quantum-ransomware-attack-d0e776bb/

 

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

 

[7] https://github.com/pan-unit42/tweets/blob/master/2022-06-28-IOCs-for-TA578-IcedID-Cobalt-Strike-and-DarkVNC.txt 

 

[8] https://github.com/stamparm/maltrail/blob/master/trails/static/malware/icedid.txt

 

[9], [15] https://www.cynet.com/blog/shelob-moonlight-spinning-a-larger-web-from-icedid-to-conti-a-trojan-and-ransomware-collaboration/

 

[10] https://www.microsoft.com/security/blog/2021/04/09/investigating-a-unique-form-of-email-delivery-for-icedid-malware/

 

[11] https://twitter.com/0xToxin/status/1564289244084011014

 

[13], [27] https://cybernews.com/security/quantum-ransomware-gang-fast-and-furious/

 

[17] https://www.virustotal.com/gui/domain/gedabuyisi.com/relations

 

[18] https://www.virustotal.com/gui/domain/sezijiru.com/relations.

 

[19] https://github.com/ByteSecLabs/ja3-ja3s-combo/blob/master/master-list.txt 

 

[21] https://www.darkreading.com/perimeter/ftp-hacking-on-the-rise

 

[22] https://www.pcrisk.com/removal-guides/23352-quantum-ransomware

 

[23] https://www.cohesity.com/resource-assets/tip-sheet/5-ways-ransomware-renders-backup-useless-tip-sheet-en.pdf

 

[24] https://www.forbes.com/sites/nishatalagala/2022/03/02/data-as-the-new-oil-is-not-enough-four-principles-for-avoiding-data-fires/ 

 

[25] https://www.bleepingcomputer.com/news/security/access-to-hacked-corporate-networks-still-strong-but-sales-fall/

 

[26] https://www.bleepingcomputer.com/news/security/ransom-payments-fall-as-fewer-victims-choose-to-pay-hackers/ 

Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Nicole Wong
Cyber Security Analyst

More in this series

No items found.

Blog

/

AI

/

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

/

AI

/

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
あなたのデータ × DarktraceのAI
唯一無二のDarktrace AIで、ネットワークセキュリティを次の次元へ