Blog
/
Network
/
August 2, 2023

Darktrace's Detection of Ransomware & Syssphinx

Read how Darktrace identified an attack technique by the threat group, Syssphinx. Learn how Darktrace's quick identification process can spot a threat.
Written by
Adam Potter
Senior Cyber Analyst
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
Adam Potter
Senior Cyber Analyst
Default blog image
02
Aug 2023

Introduction

As the threat of costly cyber-attacks continues represent a real concern to security teams across the threat landscape, more and more organizations are strengthening their defenses with additional security tools to identify attacks and protect their networks. As a result, malicious actors are being forced to adapt their tactics, modify existing variants of malicious software, or utilize entirely new variants.  

Symantec recently released an article about Syssphinx, the financially motivated cyber threat group previously known for their point-of-sale attacks. Syssphinx attempts to deploy ransomware on customer networks via a modified version of their ‘Sardonic’ backdoor. Such activity highlights the ability of threat actors to alter the composition and presentation of payloads, tools, and tactics.

Darktrace recently detected some of the same indicators suggesting a likely Syssphinx compromise within the network of a customer trialing the Darktrace DETECT™ and RESPOND™ products. Despite the potential for variations in the construction of backdoors and payloads used by the group, Darktrace’s anomaly-based approach to threat detection allowed it to stitch together a detailed account of compromise activity and identify the malicious activity prior to disruptive events on the customer’s network.

What is Syssphinx?

Syssphinx is a notorious cyber threat entity known for its financially motivated compromises.  Also referred to as FIN8, Syssphinx has been observed as early as 2016 and is largely known to target private sector entities in the retail, hospitality, insurance, IT, and financial sectors.[1]

Although Syssphinx primarily began focusing on point-of-sale style attacks, the activity associated with the group has more recently incorporated ransomware variants into their intrusions in a potential bid to further extract funds from target organizations.[2]

Syssphinx Sardonic Backdoor

Given this gradual opportunistic incorporation of ransomware, it should not be surprising that Syssphinx has slowly expanded its repertoire of tools.  When primarily performing point-of-sale compromises, the group was known for its use of point-of-sale specific malwares including BadHatch, PoSlurp/PunchTrack, and PowerSniff/PunchBuggy/ShellTea.[3]

However, in a seeming response to updates in detection systems while using previous indicators of compromise (IoCs), Syssphinx began to modify its BadHatch malware.  This resulted in the use of a C++ derived backdoor known as “Sardonic”, which has the ability to aggregate host credentials, spawn additional command sessions, and deliver payloads to compromised devices via dynamic-link library (DLL).[4],[5]

Analysis of the latest version of Sardonic reveals further changes to the malware to elude detection. These shifts include the implementation of the backdoor in the C programming language, and additional over-the-network communication obfuscation techniques. [6]

During the post-exploitation phase, the group tends to rely on “living-off-the-land” tactics, whereby an attacker utilizes tools already present within the organization’s digital environment to avoid detection. Syssphinx seems to utilize system-native tools such as PowerShell and the Windows Management Instrumentation (WMI) interface.[7] It is also not uncommon to see Windows-based vulnerability exploits employed on compromised devices. This has been observed by researchers who have examined previous iterations of Syssphinx backdoors.[8] Syssphinx also appears to exhibit elements of strategic patience and discipline in its operations, with significant time gaps in operations noted by researchers. During this time, it appears likely that updates and tweaks were applied to Syssphinx payloads.

Compromise Details

In late April 2023, Darktrace identified an active compromise on the network of a prospective customer who was trialing Darktrace DETECT+RESPOND. The customer, a retailer in EMEA with hundreds of tracked devices, reached out to the Darktrace Analyst team via the Ask the Expert (ATE) service for support and further investigation, following the encryption of their server and backup data storage in an apparent ransomware attack. Although the encryption events fell outside Darktrace’s purview due to a limited set up of trial appliances, Darktrace was able to directly track early stages of the compromise before exfiltration and encryption events began. If a full deployment had been set up and RESPOND functionality had been configured in autonomous response mode, Darktrace may have helped mitigate such encryption events and would have aided in the early identification of this ransomware attack.

Initial Intrusion and Establishment of Command and Control (C2) Infrastructure

As noted by security researchers, Syssphinx largely relies on social engineering and phishing emails to deliver its backdoor payloads. As there were no Darktrace/Email™ products deployed for this customer, it would be difficult to directly observe the exact time and manner of initial payload delivery related to this compromise. This is compounded by the fact that the customer had only recently began using Darktrace’s products during their trial period. Given the penchant for patience and delay by Syssphinx, it is possible that the intrusion began well before Darktrace had visibility of the organization’s network.

However, beginning on April 30, 2023, at 07:17:31 UTC, Darktrace observed the domain controller dc01.corp.XXXX  making repeated SSL connections to the endpoint 173-44-141-47[.]nip[.]io. In addition to the multiple open-source intelligence (OSINT) flags for this endpoint, the construction of the domain parallels that of the initial domain used to deliver a backdoor, as noted by Symantec in their analysis (37-10-71-215[.]nip[.]io). This activity likely represented the initial beaconing being performed by the compromised device. Additionally, an elevated level of incoming external data over port 443 was observed during this time, which may be associated with the delivery of the Sardonic backdoor payload. Given the unusual use of port 443 to perform SSH connections later seen in the kill chain of this attack, this activity could also parallel the employment of embedded backdoor payloads seen in the latest iteration of the Sardonic backdoor noted by Symantec.

Figure 1: Graph of the incoming external data surrounding the time of the initial establishment of command and control communication for the domain controller. As seen in the graph, the spike in incoming external data during this time may parallel the delivery of Syssphinx Sardonic backdoor.

Regardless, the domain controller proceeded to make repeated connections over port 443 to the noted domain.

Figure 2: Breach event log for the domain controller making repeated connections over port 443 to the rare external destination endpoint in constitute the establishment of C2 communication.

Internal Reconnaissance/Privilege Escalation

Following the establishment of C2 communication, Darktrace detected numerous elements of internal reconnaissance. On Apr 30, 2023, at 22:06:26 UTC, the desktop device desktop_02.corp.XXXX proceeded to perform more than 100 DRSGetNCChanges requests to the aforementioned domain controller. These commands, which are typically implemented over the RPC protocol on the DRSUAPI interface, are frequently utilized in Active Directory sync attacks to copy Active Directory information from domain controllers. Such activity, when not performed by new domain controllers to sync Active Directory contents, can indicate malicious domain or user enumeration, credential compromise or Active Directory enumeration.

Although the affected device made these requests to the previously noted domain controller, which was already compromised, such activity may have further enabled the compromise by allowing the threat actor to transfer these details to a more easily manageable device.

The device performing these DRSGetNCChanges requests would later be seen performing lateral movement activity and making connections to malicious endpoints.

Figure 3: Breach log highlighting the DRS operations performed by the corporate device to the destination domain controller. Such activity is rarely authorized for devices not tagged as administrative or as domain controllers.

Execution and Lateral Movement

At 23:09:53 UTC on April 30, 2023, the original domain server proceeded to make multiple uncommon WMI calls to a destination server on the same subnet (server01.corp.XXXX). Specifically, the device was observed making multiple RPC calls to IWbem endpoints on the server, which included login and ExecMethod (method execution) commands on the destination device. This destination device later proceeded to conduct additional beaconing activity to C2 endpoints and exfiltrate data.

Figure 4: Breach log for the domain controller performing WMI commands to the destination server during the lateral movement phase of the breach.

Similarly, beginning on May 1, 2023, at 00:11:09 UTC, the device desktop_02.corp.XXXX made multiple WMI requests to two additional devices, one server and one desktop, within the same subnet as the original domain controller. During this time, desktop_02.corp.XXXX  also utilized SMBv1, an outdated and typically non-compliant version communication protocol, to write the file rclone.exe to the same two destination devices. Rclone.exe, and its accompanying bat file, is a command-line tool developed by IT provider Rclone, to perform file management tasks. During this time, Darktrace also observed the device reading and deleting an unexpected numeric file on the ADMIN$ of the destination server, which may represent additional defense evasion techniques and tool staging.

Figure 5: Event log highlighting the writing of rclone.exe using the outdated SMBv1 communication protocol.
Figure 6: SMB logs indicating the reading and deletion of numeric string files on ADMIN$ shares of the destination devices during the time of the rclone.exe SMB writes. Such activity may be associated with tool staging and could indicate potential defense evasion techniques.

Given that the net loader sample analyzed by Symantec injects the backdoor into a WmiPrvSE.exe process, the use of WMI operations is not unexpected. Employment of WMI also correlates with the previously mentioned “living-off-the-land” tactics, as WMI services are commonly used for regular network and system administration purposes. Moreover, the staging of rclone.exe, a legitimate file management tool, for data exfiltration underscores attempts to blend into existing and expected network traffic and remain undetected on the customer’s network.

Data Exfiltration and Impact

Initial stages of data exfiltration actually began prior to some of the lateral movement events described above. On April 30, 2023, 23:09:47 the device server01.corp.XXXX, transferred nearly 11 GB of data to 173.44[.]141[.]47, as well as to the rare external IP address 170.130[.]55[.]77, which appears to have served as the main exfiltration destination during this compromise. Furthermore, the host made repeated connections to the same external IP associated with the initial suspicious beaconing activity (173.44[.]141[.]47) over SSL.

While the data exfiltration event unfolded, the device, server01.corp.XXXX, made multiple HTTP requests to 37.10[.]71[.]215, which featured URIs requesting the rclone.exe and rclone.bat files. This IP address was directly involved in the sample analyzed by Symantec. Furthermore, one of the devices that received the SMB file writes of rclone.exe and the WMI commands from desktop_02.corp.XXXX also performed SSL beaconing to endpoints associated with the compromise.

Between 01:20:45 - 03:31:41 UTC on May 1, 2023, a Darktrace detected a series of devices on the network performing a repeated pattern of activity, namely external connectivity followed by suspicious file downloads and external data transfer operations. Specifically, each affected device made multiple HTTP requests to 37.10[.]71[.]215 for rclone files. The devices proceeded to download the executable and/or binary files, and then transfer large amounts of data to the aforementioned endpoints, 170.130[.]55[.]77 and or 173-44-141-47[.]nip[.]io. Although the devices involved in data exfiltration utilized port 443 as a destination port, the connections actually used the SSH protocol. Darktrace recognized this behavior as unusual as port 443 is typically associated with the SSL protocol, while port 22 is reserved for SSH. Therefore, this activity may represent the threat actor’s attempts to remain undetected by security tools.

This unexpected use of SSH over port 443 also correlates with the descriptions of the new Sardonic backdoor according to threat researchers. Further beaconing and exfiltration activity was performed by an additional host one day later whereby the device made suspicious repeated connections to the aforementioned external hosts.

Figure 7: Connection details highlighting the use of port 443 for SSH connections during the exfiltration events.

In total, nine separate devices were involved in this pattern of activity. Five of these devices were labeled as ‘administrative’ devices according to their hostnames. Over the course of the entire exfiltration event, the attackers exfiltrated almost 61 GB of data from the organization’s environment.

Figure 8: Graph showing the levels of external data transfer from a breach device for one day on either side of the breach time. There is a large spike in such activity during the time of the breach that underscores the exfiltration events.

In addition to the individual anomaly detections by DETECT, Darktrace’s Cyber AI Analyst™ launched an autonomous investigation into the unusual behavior carried out by affected devices, connecting and collating multiple security events into one AI Analyst Incident. AI Analyst ensures that Darktrace can recognize and link the individual steps of a wider attack, rather than just identifying isolated incidents. While traditional security tools may mistake individual breaches as standalone activity, Darktrace’s AI allows it to provide unparalleled visibility over emerging attacks and their kill chains. Furthermore, Cyber AI Analyst’s instant autonomous investigations help to save customer security teams invaluable time in triaging incidents in comparison with human teams who would have to commit precious time and resources to conduct similar pattern analysis.

In this specific case, AI Analyst identified 44 separate security events from 18 different devices and was able to tie them together into one incident. The events that made up this AI Analyst Incident included:

  • Possible SSL Command and Control
  • Possible HTTP Command and Control
  • Unusual Repeated Connections
  • Suspicious Directory Replication ServiceActivity
  • Device / New or Uncommon WMI Activity
  • SMB Write of Suspicious File
  • Suspicious File Download
  • Unusual External Data Transfer
  • Unusual External Data Transfer to MultipleRelated Endpoints
Figure 9: Cyber AI Incident log highlighting multiple unusual anomalies and connecting them into one incident.

Had Darktrace RESPOND been enabled in autonomous response mode on the network of this prospective customer, it would have been able to take rapid mitigative action to block the malicious external connections used for C2 communication and subsequent data exfiltration, ideally halting the attack at this stage. As previously discussed, the limited network configuration of this trial customer meant that the encryption events unfortunately took place outside of Darktrace’s scope. When fully configured on a customer environment, Darktrace DETECT can identify such encryption attempts as soon as they occur. Darktrace RESPOND, in turn, would be able to immediately intervene by applying preventative actions like blocking internal connections that may represent file encryption, or limiting potentially compromised devices to a previously established pattern of life, ensuring they cannot carry out any suspicious activity.

Conclusion

Despite the limitations posed by the customer’s trial configuration, Darktrace demonstrated its ability to detect malicious activity associated with Syssphinx and track it across multiple stages of the kill chain.

Darktrace’s ability to identify the early stages of a compromise and various steps of the kill chain, highlights the necessity for machine learning-enabled, anomaly-based detection. In the face of threats such as Syssphinx, that exhibit the propensity to recast backdoor payloads and incorporate on “living-off-the-land” tactics, signatures and rules-based detection may not prove as effective. While Syssphinx and other threat groups will continue to adopt new tools, methods, and techniques, Darktrace’s Self-Learning AI is uniquely positioned to meet the challenge of such threats.

Appendix

DETECT Model Breaches Observed

•      Anomalous Server Activity / Anomalous External Activity from Critical Network Device

•      Anomalous Connection / Anomalous DRSGetNCChanges Operation

•      Device / New or Uncommon WMI Activity

•      Compliance / SMB Drive Write

•      Anomalous Connection / Data Sent to Rare Domain

•      Anomalous Connection / Uncommon 1 GiB Outbound

•      Unusual Activity / Unusual External Data Transfer

•      Unusual Activity / Unusual External Data to New Endpoints

•      Compliance / SSH to Rare External Destination

•      Anomalous Connection / Unusual SMB Version 1 Connectivity

•      Anomalous File / EXE from Rare External Location

•      Anomalous File / Script from Rare External Location

•      Compromise / Suspicious File and C2

•      Device / Initial Breach Chain Compromise

‍

AI Analyst Incidents Observed

•      Possible SSL Command and Control

•      Possible HTTP Command and Control

•      Unusual Repeated Connections

•      Suspicious Directory Replication Service Activity

•      Device / New or Uncommon WMI Activity

•      SMB Write of Suspicious File

•      Suspicious File Download

•      Unusual External Data Transfer

•      Unusual External Data Transfer to Multiple Related Endpoints

IoCs

IoC - Type - Description

37.10[.]71[.]215 – IP – C2 + payload endpoint

173-44-141-47[.]nip[.]io – Hostname – C2 – payload

173.44[.]141[.]47 – IP – C2 + potential payload

170.130[.]55[.]77 – IP – Data exfiltration endpoint

Rclone.exe – Exe File – Common data tool

Rclone.bat – Script file – Common data tool

MITRE ATT&CK Mapping

Command and Control

T1071 - Application Layer Protocol

T1071.001 – Web protocols

T1573 – Encrypted channels

T1573.001 – Symmetric encryption

T1573.002 – Asymmetric encryption

T1571 – Non-standard port

T1105 – Ingress tool transfer

Execution

T1047 – Windows Management Instrumentation

Credential Access‍

T1003 – OS Credential Dumping

T1003.006 – DCSync

Lateral Movement

T1570 – Lateral Tool Transfer

T1021 - Remote Services

T1021.002 - SMB/Windows Admin Shares

T1021.006 – Windows Remote Management

Exfiltration

T1048 - Exfiltration Over Alternative Protocol

T1048.001 - Exfiltration Over Symmetric Encrypted Non-C2 Protocol

T1048.002 - Exfiltration Over Symmetric Encrypted Non-C2 Protocol

T1041 - Exfiltration Over C2 Channel

References

[1] https://cyberscoop.com/syssphinx-cybercrime-ransomware/

[2] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/Syssphinx-FIN8-backdoor

[3] https://www.bleepingcomputer.com/news/security/fin8-deploys-alphv-ransomware-using-sardonic-malware-variant/

[4] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/Syssphinx-FIN8-backdoor

[5] https://thehackernews.com/2023/07/fin8-group-using-modified-sardonic.html

[6] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/Syssphinx-FIN8-backdoor

[7] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/Syssphinx-FIN8-backdoor

[8] https://www.mandiant.com/resources/blog/windows-zero-day-payment-cards

Written by
Adam Potter
Senior Cyber Analyst
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
Adam Potter
Senior Cyber Analyst

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
Network
•
August 6, 2026

When AI Agents Attack: The Case for Behavioral Anomaly Detection

A new UK AI Security Institute report shows autonomous AI agents independently developing novel attack chains, adapting when detected, and targeting humans and other AI systems. As agents become embedded across the enterprise, behavioral anomaly detection will be critical to securing the new attack surface they create.
Adam Stevens
Senior Director of Product
Read more
Network
•
July 27, 2026

Hiding in Plain Sight: Uncovering a Multi-Stage Ransomware Attack Through Behavioral Detection

Darktrace detected a multi-stage ransomware intrusion from its earliest stages, identifying reconnaissance, privilege escalation, lateral movement, command-and-control communications, and data exfiltration before encryption occurred. This analysis highlights how behavioral detection exposed attacker activity using legitimate tools and infrastructure, providing multiple opportunities for early intervention and containment.
Alexandra Evzona
Cyber Analyst
Read more
Network
•
July 1, 2026

9 Stages of Ransomware & How AI Responds

Discover the 9 stages of ransomware attacks and how AI responds at each stage. Learn how you can protect your business from cyber threats.
The Darktrace Community
Read more

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