Blog
/
Identity
/
April 29, 2025

MFA Under Attack: AiTM Phishing Kits Abusing Legitimate Services

Tycoon 2FA uses AiTM phishing and legitimate services to bypass MFA. Darktrace AI stopped it, read the blog to learn how Self-Learning AI detects sophisticated 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
fingerprintDefault blog image
29
Apr 2025

In late 2024 and early 2025, the Darktrace Security Operations Center (SOC) investigated alerts regarding separate cases of Software-as-a-Service (SaaS) account compromises on two customer environments that presented several similarities, suggesting they were part of a wider phishing campaign.

This campaign was found to leverage the project collaboration and note-taking application, Milanote, and the Tycoon 2FA phishing kit.

Legitimate services abused

As highlighted in Darktrace's 2024 Annual Threat Report [1], threat actors are abusing legitimate services, like Milanote, in their phishing campaigns. By leveraging these trusted platforms and domains, malicious actors can bypass traditional security measures, making their phishing emails appear benign and increasing the likelihood of successful attacks.

Darktrace categorizes these senders and platforms as free content senders. These services allow users to send emails containing custom content (e.g., files) from fully validated, fixed service address belonging to legitimate corporations. Although some of these services permit full body and subject customization by attackers, the structure of these emails is generally consistent, making it challenging to differentiate between legitimate and malicious emails.

What is Tycoon 2FA?

Tycoon 2FA is an Adversary-in-the-Middle (AitM) phishing kit, first seen in August 2023 and distributed via the Phishing-as-a-Service (PhaaS) model [2]. It targets multi-factor authentication (MFA) by intercepting credentials and MFA tokens during authentication on fake Microsoft or Google login pages. The attacker captures session cookies after MFA is completed, allowing them to replay the session and access the user account, even if credentials are reset. The rise in MFA use has increased the popularity of AitM phishing kits like Tycoon 2FA and Mamba 2FA, another AiTM phishing kit investigated by Darktrace.

Initial access via phishing email

At the beginning of 2025, Darktrace observed phishing emails leveraging Milanote being sent to multiple internal recipients in an organization. In this attack, the same email was sent to 19 different users, all of which were held by Darktrace.

The subject line of the emails mentioned both a legitimate internal user of the company, the company name, as well as a Milanote board regarding a “new agreement” in German. It is a common social engineering technique to mention urgent matters, such as unpaid invoices, expired passwords, or awaiting voicemails, in the subject line to prompt immediate action from the user. However, this tactic is now widely covered in phishing awareness training, making users more suspicious of such emails. In this case, while the subject mentioned a “new agreement,” likely raising the recipient’s curiosity, the tone remained professional and not overly alarming. Additionally, the mention of a colleague and the standardized language typical of free content sender emails further helped dispel concerns regarding the email.

These emails were sent by the legitimate address support@milanote[.]com and referenced "Milanote" in the personal field of the header but originated from the freemail address “ahnermatternk.ef.od.13@gmail[.]com”. Darktrace / EMAIL recognized that none of the recipients had previously received a file share email from Milanote, making this sender unfamiliar in the customer's email environment

The emails contained several benign links to legitimate Milanote endpoints (including an unsubscribe link) which were not flagged by Darktrace. However, they also included a malicious link designed to direct recipients to a pre-filled credential harvesting page hosted on Milanote, prompting them to register for an account. Despite not blocking the legitimate Milanote links in the same email, Darktrace locked the malicious link, preventing users from visiting the credential harvester.

Credential harvesting page sent to recipients, as seen in. sandbox environment.
Figure 1: Credential harvesting page sent to recipients, as seen in. sandbox environment.

Around one minute later, one recipient received a legitimate email from Milanote confirming their successful account registration, indicating they had accessed the phishing page. This email had a lower anomaly score and was not flagged by Darktrace / EMAIL because, unlike the first email, it did not contain any suspicious links and was a genuine account registration notification. Similarly, in the malicious Milanote email, only the link leading to the phishing page was blocked, while the benign and legitimate Milanote links remained accessible, demonstrating Darktrace’s precise and targeted actioning.

A legitimate and a malicious Milanote email received by one recipient.
Figure 2: A legitimate and a malicious Milanote email received by one recipient.

Around the same time, Darktrace / NETWORK observed the same user’s device making DNS query for the domain name “lrn.ialeahed[.]com” , which has been flagged as a Tycoon 2FA domain [2], suggesting the use of this phishing platform.

Once the user had entered their details in the credential harvester, it is likely that they were presented a document hosted on Milanote that contained the final payload link – likely hidden behind text instructing users to access a “new agreement” document.

External research indicates that the user was likely directed to a Cloudflare Turnstile challenge meant to reroute unwanted traffic, such as automated security scripts and penetration testing tools [2] [3]. After these checks and other background processes are completed, the user is directed to the final landing page. In this case, it was likely a fake login prompt hosted on the attacker’s server, where the user is asked to authenticate to their account using MFA. By burrowing malicious links and files in this manner, threat actors can evade analysis by traditional security email gateways, effectively bypassing their protection.

Darktrace’s analysis of the structure and word content of the phishing emails resulted in an 82% probability score that the email was malicious, and the email further received a 67% phishing inducement score, representing how closely the structure and word content of the emails compared to typical phishing emails.

All these unusual elements triggered multiple alerts in Darktrace / EMAIL, focusing on two main suspicious aspects: a new, unknown sender with no prior correspondence with the recipients or the environment, and the inclusion of a link to a previously unseen file storage solution.

Milanote phishing email as seen within Darktrace / EMAIL.
Figure 3: Milanote phishing email as seen within Darktrace / EMAIL.

After detecting the fifth email, the “Sender Surge” model alert was triggered in Darktrace / EMAIL due to a significant number of recipients being emailed by this new suspicious sender in a short period. These recipients were from various departments across the customer’s organization, including sales, marketing, purchasing, and production. Darktrace / EMAIL determined that the emails were sent to a highly unusual group of internal recipients, further raising doubts about the business legitimacy.

Darktrace / EMAIL suggested actions to contain the attack by holding all Milanote phishing emails back from recipient’s inboxes, except for the detailed email with locked links. However, autonomous actions were not enabled at the time, allowing the initial email to reach recipients' inboxes, providing a brief window for interaction. Unfortunately, during this window, one recipient clicked on the Milanote payload link, leading to the compromise of their account.

SaaS account takeover

About three minutes after the malicious Milanote email was received, Darktrace / IDENTITY detected an unusual login to the email recipient’s SaaS account. The SaaS actor was observed accessing files from their usual location in Germany, while simultaneously, a 100% rare login occurred from a location in the US that had never been seen in the customer’s environment before. This login was also flagged as suspicious by Microsoft 365, triggering a 'Conditional Access Policy' that required MFA authentication, which was successfully completed.

Tycoon 2FA adnimistration panel login page dated from October 2023 [3].
Figure 4: Tycoon 2FA adnimistration panel login page dated from October 2023 [3].

Despite the successful authentication, Darktrace / IDENTITY recognized that the login from this unusual location, coupled with simultaneous activity in another geographically distant location, were highly suspicious. Darktrace went on to observe MFA-validated logins from three separate US-based IP addresses: 89.185.80[.]19, 5.181.3[.]68, and 38.242.7[.]252. Most of the malicious activity was performed from the latter, which is associated with the Hide My Ass (HMA) VPN network [5].

Darktrace’s detection of the suspicious login from the US while the legitimate user was logged in from Germany.
Figure 5: Darktrace’s detection of the suspicious login from the US while the legitimate user was logged in from Germany.
Darktrace’s detection of the suspicious login following successful MFA authentication.
Figure 6: Darktrace’s detection of the suspicious login following successful MFA authentication.

Following this, the malicious actor accessed the user’s inbox and created a new mailbox rule named “GTH” that deleted any incoming email containing the string “milanote” in the subject line or body. Rules like this are a common technique used by attackers to leverage compromised accounts for launching phishing campaigns and concealing replies to phishing emails that might raise suspicions among legitimate account holders. Using legitimate, albeit compromised, accounts to send additional phishing emails enhances the apparent legitimacy of the malicious emails. This tactic has been reported as being used by Tycoon 2FA attackers [4].

The attacker accessed over 140 emails within the legitimate user’s inbox, including both the inbox and the “Sent Items” folder. Notably, the attacker accessed five emails in the “Sent Items” folder and modified their attachments. These emails were mainly related to invoices, suggesting the threat actor may have been looking to hijack those email threads to send fake invoices or replicate previous invoice emails.

Darktrace’s Cyber AI AnalystTM launched autonomous investigations into the individual events surrounding this suspicious activity. It connected these separate events into a single, broad account takeover incident, providing the customer with a clearer view of the ongoing compromise.

Cyber AI Analyst’s detection of unusual SaaS account activities in a single incident.
Figure 7: Cyber AI Analyst’s detection of unusual SaaS account activities in a single incident.
Cyber AI Analyst investigation of suspicious activities performed by the attacker.
Figure 8: Cyber AI Analyst investigation of suspicious activities performed by the attacker.

Darktrace's response

Within three minutes of the first unusual login alert, Darktrace’s Autonomous Response intervened, disabling the compromised user account for two hours.

As the impacted customer was subscribed to the Managed Threat Detection Service, Darktrace’s SOC team investigated the activity further and promptly alerted the customer’s security team. With the user’s account still disabled by Autonomous Response, the attack was contained, allowing the customer’s security team valuable time to investigate and remediate. Within ten minutes of receiving the alert from Darktrace’s SOC, they reset the user’s password, closed all active SaaS sessions, and deleted the malicious email rule. Darktrace’s SOC further supported the customer through the Security Operations Service Support service by providing information about the data accessed and identifying any other affected users.

Autonomous Response actions carried out by Darktrace / IDENTITY to contain the malicious activity
Figure 9: Autonomous Response actions carried out by Darktrace / IDENTITY to contain the malicious activity.

A wider Milanote phishing campaign?

Around a month before this compromise activity, Darktrace alerted another customer to similar activities involving two compromised user accounts. These accounts created new inbox rules named “GFH” and “GVB” to delete all incoming emails containing the string “milanote” in their subject line and/or body.

The phishing emails that led to the compromise of these user accounts were similar to the ones discussed above. Specifically, these emails were sent via the Milanote platform and referenced a “new agreement” (in Spanish) being shared by a colleague. Additionally, the payload link included in the phishing emails showed the same UserPrincipalName (UPN) attribute (i.e., click?upn=u001.qLX9yCzR), which has been seen in other Milanote phishing emails leveraging Tycoon 2FA reported by OSINT sources [6]. Interestingly, in some cases, the email also referenced a “new agreement” in Portuguese, indicating a global campaign.

Based on the similarities in the rule’s naming convention and action, as well as the similarities in the phishing email subjects, it is likely that these were part of the same campaign leveraging Milanote and Tycoon 2FA to compromise user accounts. Since its introduction, the Tycoon 2FA phishing kit has undergone several enhancements to increase its stealth and obfuscation methods, making it harder for security tools to detect. For example, the latest versions contain special source code to obstruct web page analysis by defenders, prevent users from copying meaningful text from the phishing webpages, and disable the right-click menu to prevent offline analysis [4].

Conclusion

Threat actors are continually employing new methods to bypass security detection tools and measures. As highlighted in this blog, even robust security mechanisms like MFA can be compromised using AitM phishing kits. The misuse of legitimate services such as Milanote for malicious purposes can help attackers evade traditional email security solutions by blurring the distinction between legitimate and malicious content.

This is why security tools based on anomaly detection are crucial for defending against such attacks. However, user awareness is equally important. Delays in processing can impact the speed of response, making it essential for users to be informed about these threats.

[related-resource]

Appendices

References

[1] https://www.darktrace.com/resources/annual-threat-report-2024

[2] https://www.validin.com/blog/tycoon_2fa_analyzing_and_hunting_phishing-as-a-service_domains

[3] https://blog.sekoia.io/tycoon-2fa-an-in-depth-analysis-of-the-latest-version-of-the-aitm-phishing-kit/#h-iocs-amp-technical-details

[4] https://blog.barracuda.com/2025/01/22/threat-spotlight-tycoon-2fa-phishing-kit

[5] https://spur.us/context/38.242.7.252    

[6] https://any.run/report/5ef1ac94e4c6c1dc35579321c206453aea80d414108f9f77abd2e2b03ffbd658/be5351d9-53c0-470b-8708-ee2e29300e70

Indicators of Compromise (IoCs)

IoC         Type      Description + Probability

89.185.80[.]19 - IP Address - Malicious login

5.181.3[.]68 - IP Address -Malicious login

38.242.7[.]252 - IP Address - Malicious login and new email inbox rule creation -  Hide My Ass VPN

lrn.ialeahed[.]com – Hostname - Likely Tycoon 2FA domain

Darktrace Model Detections

Email alerts

Platforms / Free Content Sender + High Sender Surge

Platforms / Free Content Sender + Sender Surge

Platforms / Free Content Sender + Unknown Initiator

Platforms / Free Content Sender

Platforms / Free Content Sender + First Time Recipient

Unusual / New Sender Surge

Unusual / Sender Surge

Antigena Anomaly / High Antigena Anomaly

Association / Unknown Sender

History / New Sender

Link / High Rarity Link to File Storage

Link/ Link To File Storage

Link / Link to File Storage + Unknown Sender

Link / Low Link Association

Platforms / Free Content Sender + First Time Initiator

Platforms / Free Content Sender + Unknown Initiator + Freemail

Platforms / Free Content Sender Link

Unusual / Anomalous Association

Unusual / Unlikely Recipient Association

IDENTITY

SaaS / Access / Unusual External Source for SaaS Credential Use

SaaS / Compromise / Login from Rare High Risk Endpoint

SaaS / Access / M365 High Risk Level Login

SaaS / Compromise / Login From Rare Endpoint While User Is Active

SaaS / Access / MailItemsAccessed from Rare Endpoint

SaaS / Unusual Activity / Multiple Unusual SaaS Activities

SaaS / Unusual Activity / Multiple Unusual External Sources For SaaS Credential

SaaS / Compliance / Anomalous New Email Rule

SaaS / Compromise / Unusual Login and New Email Rule

SaaS / Compromise / SaaS Anomaly Following Anomalous Login

Antigena / SaaS / Antigena Suspicious SaaS Activity Block

Antigena / SaaS / Antigena Enhanced Monitoring from SaaS User Block

Antigena / SaaS / Antigena Unusual Activity Block

Antigena / SaaS / Antigena Suspicious SaaS and Email Activity Block

Cyber AI Analyst Incident

Possible Hijack of Office365 Account

MITRE ATT&CK Mapping

Tactic – Technique

DEFENSE EVASION, PERSISTENCE, PRIVILEGE ESCALATION, INITIAL ACCESS - Cloud Accounts

INITIAL ACCESS - Phishing

CREDENTIAL ACCESS - Steal Web Session Cookie

PERSISTENCE - Account Manipulation

PERSISTENCE - Outlook Rules

RESOURCE DEVELOPMENT - Email Accounts

RESOURCE DEVELOPMENT - Compromise Accounts

Experts breakdown of Identity Security in the Cloud Era

This white paper discusses the current identity threat landscape and how defenders can adopt new tools to better secure their users and data.

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