Blog
/
Network
/
June 19, 2023

Darktrace Detection of 3CX Supply Chain Attack

Explore how the 3CX supply chain compromise was uncovered, revealing key insights into the detection of sophisticated cyber 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
Nahisha Nobregas
SOC Analyst
Default blog image
19
Jun 2023

Ever since the discovery of the SolarWinds hack that affected tens of thousands of organizations around the world in 2020, supply chain compromises have remained at the forefront of the minds of security teams and continue to pose a significant threat to their business operations. 

Supply chain compromises can have far-reaching implications, from disrupting an organization’s daily operations, incurring huge financial and reputational damage, to affecting the critical infrastructure of entire countries. As such, it is essential for organizations to have effective security measures in place able to identify and halt these attacks at the earliest possible stage.

In March 2023 the 3CX Desktop application became the latest victim of a supply chain compromise dubbed as the “SmoothOperator” by SentinelOne. This application is used by over 600,000 companies worldwide and the customer list contains high-profile customers across a variety of industries [2]. The 3CX Desktop application is a Voice over Internet Protocol (VoIP) communication software for enterprises that allows for chats, video calls, and voice calls. [3] The 3CX installers for both Windows and macOS systems were affected by information stealing malware. Researchers were able to discern that threat actors also known as UNC 4736 related to financially motivated North Korean operators also known as AppleJeus were responsible for the supply chain compromise.  Researchers have also linked it to another supply chain compromise that occurred prior on the Trading Technologies X_TRADER platform, making this the first known cascading software supply chain compromise used to distribute malware on a wide scale and still be able to align operator interests. [3] Customer reports following the compromise began to surface about the 3CX software being picked up as malicious by several cybersecurity vendors such as CrowdStrike, SentinelOne, and Palo Alto Networks. [6] 

By leveraging integrations with other security vendors like CrowdStrike and SentinelOne, Darktrace DETECT™ was able to identify activity from the “SmoothOperator” across the customer base at multiple stages of the kill chain in March 2023. Darktrace RESPOND™ was then able to autonomously intervene against these emerging threats, preventing significant disruption to customer networks. 

Background on the first known cascading supply chain attack 

Initial Access

In April 2023, security researchers identified the initial target in this story was not the 3CX desktop application, rather, it was another software application called X_TRADER by Trading Technologies. [3] Trading Technologies is a provider that offers high-performance financial trading packages, allowing financial professionals to analyze and trade assets within the stock market more efficiently. Unfortunately, a compromise already existed in the supply chain for this organization. The X_TRADER installer, which had been retired in 2020, still had its code signing certificate set to expire in October 2022. This code signing certificate was exploited by attackers to digitally sign the malicious software. [3] It also inopportunely led to 3CX when an employee unknowingly downloaded a trojanized installer for the X_TRADER software from Trading Technologies prior to the certificate’s expiration. [4]. This compromise of 3CX via X_TRADER was the first case of a cascading supply chain attack reported on within the wider threat landscape. 

Persistence and Privilege Escalation 

Following these findings, researchers were able to identify the likely kill chain that occurred on Windows systems, beginning with the download of the 3CX DesktopApp installer that executed an executable (.exe) file before dropping two trojanized Data Link Libraries (DLLs) alongside a benign executable that was used to sideload malicious DLLs. These DLLs contained and used SIGFLIP and DAVESHELL; both publicly available projects. [3] In this case, the DLLs were used to decrypt using an RC4 key and load a payload into the memory of a compromised system. [3] SIGFLIP and DAVESHELL also extract and decrypt the modular backdoor named VEILEDSIGNAL, which also contains a command and control (C2) configuration. This malware allowed the North Korean threat operators to gain administrative control to the 3CX employee’s device. [3] This was followed by access to the employee’s corporate credentials, ultimately leading to access to 3CX systems. [4] 

Lateral Movement and C2 activity

Security researchers were also able to identify other malware families that were mainly utilized in the supply chain attack to move laterally within the 3CX environment, and allow for C2 communication [3], these malware families are detailed below:

  • TaxHaul: when executed it decrypts shellcode payload, observed by Mandiant to persist via DLL search-order hijacking.
  • Coldcat: complex downloader, which also beacons to a C2 infrastructure.
  • PoolRat: collects system information and executes commands. This is the malware that was found to affect macOS systems.
  • IconicStealer: served as a third stage payload on 3CX systems to steal data or information.

Furthermore, it was also reported early on by Kaspersky that a backdoor named Gopuram, routinely used by the North Korean threat actors Lazarus and typically used against cryptocurrency companies, was also used as a second stage payload on a limited number of 3CX’s customers compromised systems. [5]

3CX detections observed by Darktrace

CrowdStrike and SentinelOne, two of the major detection platforms with which Darktrace partners through security integrations, initially revealed that their platforms had identified the campaign appeared to be targeting 3CXDesktopApp customers in March 2023. 

At this time, Darktrace was also observing this activity and alerting customers to unusual behavior on their networks. [1][7] Darktrace DETECT identified activity related to the supply chain compromise primarily through host-level alerts associated with CrowdStrike and SentinelOne integrations, as well as model breaches related to lateral movement and C2 activity. 

Some of the activity related to the 3CX supply chain compromise that Darktrace detected was observed solely via integration models picking up executable and Microsoft Software Installer (msi) file downloads for the 3CXDesktopApp, suggesting the compromise likely was stopped at the endpoint device. 

CrowdStrike integration model breach identifying 3CXDesktopApp[.]exe as possible malware
Figure 1: CrowdStrike integration model breach identifying 3CXDesktopApp[.]exe as possible malware on March 30, 2023.
showcases the Model Breach Event Log for the CrowdStrike integration model breach
Figure 2: The above figure, showcases the Model Breach Event Log for the CrowdStrike integration model breach shown in Figure 1.

In another case highlighted in Figure 3 and 4, security platforms were associating 3CX as malicious. The device in these figures was observed downloading a 3CXDesktopApp executable followed by an msi file about an hour later. This pattern of activity correlates with the compromise process that had been on reported, where the “SmoothOperator” malware that affected 3CX systems was able to persist through DLL side-loading of malicious DLL files delivered with benign executable files, making it difficult for traditional security tools to detect. [2][3][7]

The activity in this case was detected by the DETECT integration model, ‘High Severity Integration Malware Detection’ and was later blocked by the Darktrace RESPOND/Network model, ‘Antigena Significant Anomaly from Client Block’ which applied the “Enforce Pattern of Life” action to intercept the malicious download that was taking place. Darktrace RESPOND uses AI to learn every devices normal pattern of life and act autonomously to enforce its normal activity. In this event, RESPOND would not only intercept the malicious download that was taking place on the device, but also not allow the device to significantly deviate from its normal pattern of activity.

The Model Breach Event log for the device displays the moment in which the SentinelOne integration model breached for the 3CXDesktopApp.exe file
Figure 3: The Model Breach Event log for the device displays the moment in which the SentinelOne integration model breached for the 3CXDesktopApp.exe file followed subsequently by the RESPOND model, ‘Antigena Significant Anomaly from Client Block’, on March 29, 2023.
Another ‘High Severity Integration Malware Detection’ breached
Figure 4: Another ‘High Severity Integration Malware Detection’ breached for the same device in Figure 3 approximately one hour later because of the msi file, 3CXDesktopApp-18.12.416.msi, which also led to the Darktrace RESPOND model, ‘Antigena Significant Anomaly from Client Block’, on March 29, 2023.

In a separate case, Darktrace also detected a device performing unusual SMB drive writes for the file ‘3CXDesktopApp-18.10.461.msi’. This breached the DETECT model ‘SMB Drive Write’. This model detects when a device starts writing files to another internal device it does not usually communicate with via the SMB protocol using the admin$ or drive shares.

This Model Breach Event log highlights the moment Darktrace captured the msi application file for the 3CXDesktopApp being transferred internally on this customer’s network
Figure 5: This Model Breach Event log highlights the moment Darktrace captured the msi application file for the 3CXDesktopApp being transferred internally on this customer’s network, this was picked up as new activity for the device on March 28, 2023. 

In a couple of other cases observed by Darktrace, connections detected were made from affected devices to 3CX compromise related endpoints. In Figure 6, the device in question was detected connecting to the endpoint, journalide[.]org. This breached the model, ‘Suspicious Self-Signed SSL’, which looks for connections being made to an endpoint with a self-signed SSL certificate which is designed to look legitimate, as self-signed certificates are often used in malware communication.

Model Breach Event log for connections to the 3CX C2 related endpoint
Figure 6: Model Breach Event log for connections to the 3CX C2 related endpoint, journalide[.]org, these connections breached the model Suspicious Self-Signed SSL on April 24, 2023.

On another Darktrace customer environment, a 3CX C2 endpoint, pbxphonenetwork[.]com, had already been added to the Watched Domains list around the time reports of the 3CX application software being malicious had been reported. The Watched Domains list allows Darktrace to detect if any device on the network makes connections to these domains with more scrutiny and breach a model for further visibility of threats on the network. Activity in this case was detected and subsequently blocked by a Darktrace RESPOND action, “Block connections to 89.45.67[.]160 port 443 and pbxphonenetwork[.]com on port 443”, blocking the device from connecting to this 3CX C2 endpoints on the spot (see Figure 7). This activity subsequently breached the RESPOND model, ‘Antigena Watched Domain Block’. 

Figure 7: History log of the Darktrace RESPOND action applied to the device breaching the Darktrace RESPOND model, Antigena Watched Domain Block and applying the action, “Block connections to 89.45.67[.]160 port 443 and pbxphonenetwork[.]com on port 443” on March 31, 2023.

Darktrace Coverage 

Utilizing integrations with Darktrace such as those with CrowdStrike and SentinelOne, Darktrace was able to detect and respond to activity identified as malicious 3CX activity by CrowdStrike and SentinelOne as seen in Figures 1, 2, 3, and 4. This activity breached the following Darktrace DETECT models: 

  • Integration / CrowdStrike Alert
  • Security Integration / High Severity Integration Malware Detection

Darktrace was also able to identify lateral movement activity such as in the case illustrated in Figure 5.

  • Compliance / SMB Drive Write

Lastly, C2 beaconing activity from malicious endpoints associated with the 3CX compromise was also detected as seen in Figure 6, this activity breached the following Darktrace DETECT model:

  • Anomalous Connection / Suspicious Self-Signed SSL

For customers with Darktrace RESPOND configured in autonomous response mode, Darktrace RESPOND models also breached to activity related to the 3CX supply chain compromise as seen in Figures 3, 4, and 7. Below are the models that breached and the following autonomous actions that were applied:

  • Antigena / Network / Significant Anomaly / Antigena Significant Anomaly from Client Block, “Enforce pattern of life”
  • Antigena / Network / External Threat / Antigena Watched Domain Block, “Block connections to 89.45.67[.]160 port 443 and pbxphonenetwork[.]com on port 443”

Conclusion 

The first known cascading supply chain compromise occurred inopportunely for 3CX but conveniently for UNC 4736 North Korean threat actors. This “SmoothOperator” compromise was detected by endpoint security platforms such as CrowdStrike who was at the cusp of this discovery when it became one of the first platforms to report on malicious activity related to the 3CX DesktopApp supply chain compromise.  

Although still novel at the time and largely without reported indicators of compromise, Darktrace was able to capture and identify activity related to the 3CX compromise across its customer base, as well as respond autonomously to contain it. Darktrace was able to amplify security integrations with CrowdStrike and SentinelOne, and via anomaly-based model breaches, contribute unique insights by highlighting activity in varied parts of the 3CX supply chain compromise kill chain. The “SmoothOperator” supply chain attack proves that the Darktrace suite of products, including DETECT and RESPOND, can not only act autonomously to identify and respond to novel threats, but also work with security integrations to further amplify intervention and prevent cyber disruption on customer networks. 

Credit to Nahisha Nobregas, SOC Analyst and Trent Kessler, SOC Analyst.

Appendices

MITRE ATT&CK Framework

Resource Development

  • T1588 Obtain Capabilities  
  • T1588.004 Digital Certificates
  • T1608 Stage Capabilities  
  • T1608.003 Install Digital Certificate

Initial Access

  • T1190 Exploit Public-Facing Application
  • T1195 Supply Chain Compromise  
  • T1195.002 Compromise Software Supply Chain

Persistence

  • T1574 Hijack Execution Flow
  • T1574.002 DLL Side-Loading

Privilege Escalation

  • T1055 Process Injection
  • T1574 Hijack Execution Flow  
  • T1574.002 DLL Side-Loading

Command and Control

  • T1071 Application Layer Protocol
  • T1071.001 Web Protocols
  • T1071.004 DNS  
  • T1105 Ingress Tool Transfer
  • T1573 Encrypted Channel

List of IOCs

C2 Hostnames

  • journalide[.]org
  • pbxphonenetwork[.]com

Likely C2 IP address

  • 89.45.67[.]160

References

  1. https://www.crowdstrike.com/blog/crowdstrike-detects-and-prevents-active-intrusion-campaign-targeting-3cxdesktopapp-customers/
  2. https://www.bleepingcomputer.com/news/security/3cx-confirms-north-korean-hackers-behind-supply-chain-attack/
  3. https://www.mandiant.com/resources/blog/3cx-software-supply-chain-compromise
  4. https://www.securityweek.com/cascading-supply-chain-attack-3cx-hacked-after-employee-downloaded-trojanized-app/
  5. https://securelist.com/gopuram-backdoor-deployed-through-3cx-supply-chain-attack/109344/
  6. https://www.bleepingcomputer.com/news/security/3cx-hack-caused-by-trading-software-supply-chain-attack/
  7. https://www.sentinelone.com/blog/smoothoperator-ongoing-campaign-trojanizes-3cx-software-in-software-supply-chain-attack/
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
Nahisha Nobregas
SOC 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