Blog
/
/
April 4, 2022

Explore Internet-Facing System Vulnerabilities

Read about 2021's top four incidents and how Darktrace's advanced threat detection technology identified and mitigated vulnerabilities. Learn more.
Written by
Sam Lister
Specialist Security Researcher
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
Apr 2022

By virtue of their exposure, Internet-facing systems (i.e., systems which have ports open/exposed to the wider Internet) are particularly susceptible to compromise. Attackers typically compromise Internet-facing systems by exploiting zero-day vulnerabilities in applications they run. During 2021, critical zero-day vulnerabilities in the following applications were publicly disclosed:

Internet-facing systems running these applications were consequently heavily targeted by attackers. In this post, we will provide examples of compromises of these systems observed by Darktrace’s SOC team in 2021. As will become clear, successful exploitation of weaknesses in Internet-facing systems inevitably results in such systems doing things which they do not normally do. Rather than focusing on identifying attempts to exploit these weaknesses, Darktrace focuses on identifying the unusual behaviors which inevitably ensue. The purpose of this post is to highlight the effectiveness of this approach.

Exchange server compromise

In January, researchers from the cyber security company DEVCORE reported a series of critical vulnerabilities in Microsoft Exchange which they dubbed ‘ProxyLogon’.[1] ProxyLogon consists of a server-side request forgery (SSRF) vulnerability (CVE-2021-26855) and a remote code execution (RCE) vulnerability (CVE-2021-27065). Attackers were observed exploiting these vulnerabilities in the wild from as early as January 6.[2] In April, DEVCORE researchers reported another series of critical vulnerabilities in Microsoft Exchange which they dubbed ‘ProxyShell’.[3] ProxyShell consists of a pre-authentication path confusion vulnerability (CVE-2021-34473), a privilege elevation vulnerability (CVE-2021-34523), and a post-authentication RCE vulnerability (CVE-2021-31207). Attackers were first observed exploiting these vulnerabilities in the wild in August.[4] In many cases, attackers exploited the ProxyShell and ProxyLogon vulnerabilities in order to create web shells on the targeted Exchange servers. The presence of these web shells provided attackers with the means to remotely execute commands on the compromised servers.

In early August 2021, by exploiting the ProxyShell vulnerabilities, an attacker gained the rights to remotely execute PowerShell commands on an Internet-facing Exchange server within the network of a US-based transportation company. The attacker subsequently executed a number of PowerShell commands on the server. One of these commands caused the server to make a 28-minute-long SSL connection to a highly unusual external endpoint. Within a couple of hours, the attacker managed to strengthen their foothold within the network by installing AnyDesk and CobaltStrike on several internal devices. In mid-August, the attacker got the devices on which they had installed Cobalt Strike to conduct network reconnaissance and to transfer terabytes of data to the cloud storage service, MEGA. At the end of August, the attacker got the devices on which they had installed AnyDesk to execute Conti ransomware and to spread executable files and script files to further internal devices.

In this example, the attacker’s exploitation of ProxyShell immediately resulted in the Exchange Server making a long SSL connection to an unusual external endpoint. This connection caused the model Device / Long Agent Connection to New Endpoint to breach. The subsequent reconnaissance, lateral movement, C2, external data transfer, and encryption behavior brought about by the attacker were also picked up by Darktrace’s models.

A non-exhaustive list of the models that breached as a result of the behavior brought about by the attacker:

  • Device / Long Agent Connection to New Endpoint
  • Device / ICMP Address Scan
  • Anomalous Connection / SMB Enumeration
  • Anomalous Server Activity / Outgoing from Server
  • Compromise / Beacon to Young Endpoint
  • Anomalous Server Activity / Rare External from Server
  • Compromise / Fast Beaconing to DGA
  • Compromise / SSL or HTTP Beacon
  • Compromise / Sustained SSL or HTTP Increase
  • Compromise / Beacon for 4 Days
  • Anomalous Connection / Multiple HTTP POSTs to Rare Hostname
  • Unusual Activity / Enhanced Unusual External Data Transfer
  • Anomalous Connection / Data Sent to Rare Domain
  • Anomalous Connection / Uncommon 1 GiB Outbound
  • Compliance / SMB Drive Write
  • Anomalous File / Internal / Additional Extension Appended to SMB File
  • Anomalous Connection / Suspicious Read Write Ratio
  • Anomalous Connection / Suspicious Read Write Ratio and Unusual SMB
  • Anomalous Connection / Sustained MIME Type Conversion
  • Unusual Activity / Anomalous SMB Move & Write
  • Unusual Activity / Unusual Internal Data Volume as Client or Server
  • Device / Suspicious File Writes to Multiple Hidden SMB Shares
  • Compromise / Ransomware / Suspicious SMB Activity
  • Anomalous File / Internal / Unusual SMB Script Write
  • Anomalous File / Internal / Masqueraded Executable SMB Write
  • Device / SMB Lateral Movement
  • Device / Multiple Lateral Movement Model Breaches

Confluence server compromise

Atlassian’s Confluence is an application which provides the means for building collaborative, virtual workspaces. In the era of remote working, the value of such an application is undeniable. The public disclosure of a critical remote code execution (RCE) vulnerability (CVE-2021-26084) in Confluence in August 2021 thus provided a prime opportunity for attackers to cause havoc. The vulnerability, which arises from the use of Object-Graph Navigation Language (OGNL) in Confluence’s tag system, provides attackers with the means to remotely execute code on vulnerable Confluence server by sending a crafted HTTP request containing a malicious parameter.[5] Attackers were first observed exploiting this vulnerability towards the end of August, and in the majority of cases, attackers exploited the vulnerability in order to install crypto-mining tools onto vulnerable servers.[6]

At the beginning of September 2021, an attacker was observed exploiting CVE-2021-26084 in order to install the crypto-mining tool, XMRig, as well as a shell script, onto an Internet-facing Confluence server within the network of an EMEA-based television and broadcasting company. Within a couple of hours, the attacker installed files associated with the crypto-mining malware, Kinsing, onto the server. The Kinsing-infected server then immediately began to communicate over HTTP with the attacker’s C2 infrastructure. Around the time of this activity, the server was observed using the MinerGate crypto-mining protocol, indicating that the server had begun to mine cryptocurrency.

In this example, the attacker’s exploitation of CVE-2021-26084 immediately resulted in the Confluence server making an HTTP GET request with an unusual user-agent string (one associated with curl in this case) to a rare external IP. This behavior caused the models Device / New User Agent, Anomalous Connection / New User Agent to IP Without Hostname, and Anomalous File / Script from Rare Location to breach. The subsequent file downloads, C2 traffic and crypto-mining activity also resulted in several models breaching.

A non-exhaustive list of the models which breached as a result of the unusual behavior brought about by the attacker:

  • Device / New User Agent
  • Anomalous Connection / New User Agent to IP Without Hostname
  • Anomalous File / Script from Rare Location
  • Anomalous File / EXE from Rare External Location
  • Anomalous File / Internet Facing System File Download
  • Device / Initial Breach Chain Compromise
  • Anomalous Connection / Posting HTTP to IP Without Hostname
  • Compliance / Crypto Currency Mining Activity
  • Compromise / High Priority Crypto Currency Mining
  • Device / Internet Facing Device with High Priority Alert

GitLab server compromise

GitLab is an application providing services ranging from project planning to source code management. Back in April 2021, a critical RCE vulnerability (CVE-2021-22205) in GitLab was publicly reported by a cyber security researcher via the bug bounty platform, HackerOne.[7] The vulnerability, which arises from GitLab’s use of ExifTool for removing metadata from image files, [8] enables attackers to remotely execute code on vulnerable GitLab servers by uploading specially crafted image files.[9] Attackers were first observed exploiting CVE-2021-22205 in the wild in June/July.[10] A surge in exploitations of the vulnerability was observed at the end of October, with attackers exploiting the flaw in order to assemble botnets.[11] Darktrace observed a significant number of cases in which attackers exploited the vulnerability in order to install crypto-mining tools onto vulnerable GitLab servers.

On October 29, an attacker successfully exploited CVE-2021-22205 on an Internet-facing GitLab server within the network of a UK-based education provider. The organization was trialing Darktrace when this incident occurred. The attacker installed several executable files and shell scripts onto the server by exploiting the vulnerability. The attacker communicated with the compromised server (using unusual ports) for several days, before making the server transfer large volumes of data externally and download the crypto-mining tool, XMRig, as well as the botnet malware, Mirai. The server was consequently observed making connections to the crypto-mining pool, C3Pool.

In this example, the attacker’s exploitation of the vulnerability in GitLab immediately resulted in the server making an HTTP GET request with an unusual user-agent string (one associated with Wget in this case) to a rare external IP. The models Anomalous Connection / New User Agent to IP Without Hostname and Anomalous File / EXE from Rare External Location breached as a result of this behavior. The attacker’s subsequent activity on the server over the next few days resulted in frequent model breaches.

A non-exhaustive list of the models which breached as a result of the attacker’s activity on the server:

  • Anomalous Connection / New User Agent to IP Without Hostname
  • Anomalous File / EXE from Rare External Location
  • Anomalous File / Multiple EXE from Rare External Locations
  • Anomalous File / Internet Facing Device with High Priority Alert
  • Anomalous File / Script from Rare Location
  • Anomalous Connection / Application Protocol on Uncommon Port
  • Anomalous Connection / Anomalous SSL without SNI to New External
  • Device / Initial Breach Chain Compromise
  • Unusual Activity / Unusual External Data to New IPs
  • Anomalous Server Activity / Outgoing from Server
  • Device / Large Number of Model Breaches from Critical Network Device
  • Anomalous Connection / Data Sent to Rare Domain
  • Compromise / Suspicious File and C2
  • Unusual Activity / Enhanced Unusual External Data Transfer
  • Compliance / Crypto Currency Mining Activity
  • Compliance / High Priority Crypto Currency Mining
  • Anomalous File / Zip or Gzip from Rare External Location
  • Compromise / Monero Mining
  • Device / Internet Facing Device with High Priority Alert
  • Anomalous Server Activity / Rare External from Server
  • Compromise / Slow Beaconing Activity To External Rare
  • Compromise / Beaconing Activity To External Rare
  • Compromise / HTTP Beaconing to Rare Destination
  • Compromise / High Volume of Connections with Beacon Score
  • Anomalous File / Numeric Exe Download

Log4j server compromise

On December 9 2021, a critical RCE vulnerability (dubbed ‘Log4Shell’) in version 2 of Apache’s Log4j was publicly disclosed by researchers at LunaSec.[12] As a logging library present in potentially millions of Java applications,[13] Log4j constitutes an obscured, yet ubiquitous feature of the digital world. The vulnerability (CVE-2021-44228), which arises from Log4j’s Java Naming and Directory Interface (JNDI) Lookup feature, enables an attacker to make a vulnerable server download and execute a malicious Java class file. To exploit the vulnerability, all the attacker must do is submit a specially crafted JNDI lookup request to the server. The fact that Log4j is present in so many applications and that the exploitation of this vulnerability is so simple, Log4Shell has been dubbed the ‘most critical vulnerability of the last decade’.[14] Attackers have been exploiting Log4Shell in the wild since at least December 1.[15] Since then, attackers have been observed exploiting the vulnerability to install crypto-mining tools, Cobalt Strike, and RATs onto vulnerable servers.[16]

On December 10, one day after the public disclosure of Log4Shell, an attacker successfully exploited the vulnerability on a vulnerable Internet-facing server within the network of a US-based architecture company. By exploiting the vulnerability, the attacker managed to get the server to download and execute a Java class file named ‘Exploit69ogQNSQYz.class’. Executing the code in this file caused the server to download a shell script file and a file related to the Kinsing crypto-mining malware. The Kinsing-infected server then went on to communicate over HTTP with a C2 server. Since the customer was using the Proactive Threat Notification (PTN) service, they were immediately alerted to this activity, and the server was subsequently quarantined, preventing crypto-mining activity from taking place.

In this example, the attacker’s exploitation of the zero-day vulnerability immediately resulted in the vulnerable server making an HTTP GET request with an unusual user-agent string (one associated with Java in this case) to a rare external IP. The models Anomalous Connection / Callback on Web Facing Device and Anomalous Connection / New User Agent to IP Without Hostname breached as a result of this behavior. The device’s subsequent file downloads and C2 activity caused several Darktrace models to breach.

A non-exhaustive list of the models which breached as a result of the unusual behavior brought about by the attacker:

  • Anomalous Connection / Callback on Web Facing Device
  • Anomalous Connection / New User Agent to IP Without Hostname
  • Anomalous File / Internet Facing System File Download
  • Anomalous File / Script from Rare External Location
  • Device / Initial Breach Chain Compromise
  • Anomalous Connection / Posting HTTP to IP Without Hostname

Round-up

It is inevitable that attackers will attempt to exploit zero-day vulnerabilities in applications running on Internet-facing devices. Whilst identifying these attempts is useful, the fact that attackers regularly exploit new zero-days makes the task of identifying attempts to exploit them akin to a game of whack-a-mole. Whilst it is uncertain which zero-day vulnerability attackers will exploit next, what is certain is that their exploitation of it will bring about unusual behavior. No matter the vulnerability, whether it be a vulnerability in Microsoft Exchange, Confluence, GitLab, or Log4j, Darktrace will identify the unusual behaviors which inevitably result from its exploitation. By identifying unusual behaviors displayed by Internet-facing devices, Darktrace thus makes it almost impossible for attackers to successfully exploit zero-day vulnerabilities without being detected.

For Darktrace customers who want to find out more about detecting potential compromises of internet-facing devices, refer here for an exclusive supplement to this blog.

Thanks to Andy Lawrence for his contributions.

Footnotes

1. https://devco.re/blog/2021/08/06/a-new-attack-surface-on-MS-exchange-part-1-ProxyLogon/

2. https://www.volexity.com/blog/2021/03/02/active-exploitation-of-microsoft-exchange-zero-day-vulnerabilities/

3. https://www.zerodayinitiative.com/blog/2021/8/17/from-pwn2own-2021-a-new-attack-surface-on-microsoft-exchange-proxyshell

4. https://www.rapid7.com/blog/post/2021/08/12/proxyshell-more-widespread-exploitation-of-microsoft-exchange-servers/

5. https://www.kaspersky.co.uk/blog/confluence-server-cve-2021-26084/23376/

6. https://www.bleepingcomputer.com/news/security/atlassian-confluence-flaw-actively-exploited-to-install-cryptominers/

7. https://hackerone.com/reports/1154542

8. https://security.humanativaspa.it/gitlab-ce-cve-2021-22205-in-the-wild/

9.https://about.gitlab.com/releases/2021/04/14/security-release-gitlab-13-10-3-released/

10. https://www.rapid7.com/blog/post/2021/11/01/gitlab-unauthenticated-remote-code-execution-cve-2021-22205-exploited-in-the-wild/

11. https://www.hackmageddon.com/2021/12/16/1-15-november-2021-cyber-attacks-timeline/

12. https://www.lunasec.io/docs/blog/log4j-zero-day/

13. https://www.csoonline.com/article/3644472/apache-log4j-vulnerability-actively-exploited-impacting-millions-of-java-based-apps.html

14. https://www.theguardian.com/technology/2021/dec/10/software-flaw-most-critical-vulnerability-log-4-shell

15. https://www.rapid7.com/blog/post/2021/12/15/the-everypersons-guide-to-log4shell-cve-2021-44228/

16. https://www.microsoft.com/security/blog/2021/12/11/guidance-for-preventing-detecting-and-hunting-for-cve-2021-44228-log4j-2-exploitation/

Written by
Sam Lister
Specialist Security Researcher
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

Botnet Behind the Camera: Mirai Katana Activity on a Video Recording Device

OT
•
September 2, 2026
Parvatha Ananthakannan
Cyber Analyst
Watch the NIS2 Webinar
Trending blogs
1
Prompt Security in Enterprise AI: Strengths, Weaknesses, and Common Approaches
May 20, 2026
2
How Empowering End Users can Improve Your Email Security and Decrease the Burden on the SOC
May 8, 2024
3
Elevating Network Security: Confronting Trust, Ransomware, & Novel Attacks
Jun 21, 2024
4
The State of AI in Cybersecurity: Unveiling Global Insights from 1,800 Security Practitioners
Apr 9, 2024
5
The State of AI in Cybersecurity: The Impact of AI on Cybersecurity Solutions
May 13, 2024

More in this series

No items found.
Continue reading
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