Blog
/
Network
/
October 9, 2025

Inside Akira’s SonicWall Campaign: Darktrace’s Detection and Response

Starting in July 2025, Akira ransomware attacks surged globally, targeting SonicWall SSL VPN devices. In August, Darktrace detected suspicious activity in a US network, including scanning, lateral movement, and data exfiltration. A compromised SonicWall VPN server linked the incident to the broader Akira campaign exploiting known vulnerabilities.
Written by
Emily Megan Lim
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
Emily Megan Lim
Cyber Analyst
akira sonicwallDefault blog image
09
Oct 2025

Introduction: Background on Akira SonicWall campaign

Between July and August 2025, security teams worldwide observed a surge in Akira ransomware incidents involving SonicWall SSL VPN devices [1]. Initially believed to be the result of an unknown zero-day vulnerability, SonicWall later released an advisory announcing that the activity was strongly linked to a previously disclosed vulnerability, CVE-2024-40766, first identified over a year earlier [2].

On August 20, 2025, Darktrace observed unusual activity on the network of a customer in the US. Darktrace detected a range of suspicious activity, including network scanning and reconnaissance, lateral movement, privilege escalation, and data exfiltration. One of the compromised devices was later identified as a SonicWall virtual private network (VPN) server, suggesting that the incident was part of the broader Akira ransomware campaign targeting SonicWall technology.

As the customer was subscribed to the Managed Detection and Response (MDR) service, Darktrace’s Security Operations Centre (SOC) team was able to rapidly triage critical alerts, restrict the activity of affected devices, and notify the customer of the threat. As a result, the impact of the attack was limited - approximately 2 GiB of data had been observed leaving the network, but any further escalation of malicious activity was stopped.

Threat Overview

CVE-2024-40766 and other misconfigurations

CVE-2024-40766 is an improper access control vulnerability in SonicWall’s SonicOS, affecting Gen 5, Gen 6, and Gen 7 devices running SonicOS version 7.0.1 5035 and earlier [3]. The vulnerability was disclosed on August 23, 2024, with a patch released the same day. Shortly after, it was reported to be exploited in the wild by Akira ransomware affiliates and others [4].

Almost a year later, the same vulnerability is being actively targeted again by the Akira ransomware group. In addition to exploiting unpatched devices affected by CVE-2024-40766, security researchers have identified three other risks potentially being leveraged by the group [5]:

*The Virtual Office Portal can be used to initially set up MFA/TOTP configurations for SSLVPN users.

Thus, even if SonicWall devices were patched, threat actors could still target them for initial access by reusing previously stolen credentials and exploiting other misconfigurations.

Akira Ransomware

Akira ransomware was first observed in the wild in March 2023 and has since become one of the most prolific ransomware strains across the threat landscape [6]. The group operates under a Ransomware-as-a-Service (RaaS) model and frequently uses double extortion tactics, pressuring victims to pay not only to decrypt files but also to prevent the public release of sensitive exfiltrated data.

The ransomware initially targeted Windows systems, but a Linux variant was later observed targeting VMware ESXi virtual machines [7]. In 2024, it was assessed that Akira would continue to target ESXi hypervisors, making attacks highly disruptive due to the central role of virtualisation in large-scale cloud deployments. Encrypting the ESXi file system enables rapid and widespread encryption with minimal lateral movement or credential theft. The lack of comprehensive security protections on many ESXi hypervisors also makes them an attractive target for ransomware operators [8].

Victimology

Akira is known to target organizations across multiple sectors, most notably those in manufacturing, education, and healthcare. These targets span multiple geographic regions, including North America, Latin America, Europe and Asia-Pacific [9].

Geographical distribution of organization’s affected by Akira ransomware in 2025 [9].
Figure 1: Geographical distribution of organization’s affected by Akira ransomware in 2025 [9].

Common Tactics, Techniques and Procedures (TTPs) [7][10]

Initial Access
Targets remote access services such as RDP and VPN through vulnerability exploitation or stolen credentials.

Reconnaissance
Uses network scanning tools like SoftPerfect and Advanced IP Scanner to map the environment and identify targets.

Lateral Movement
Moves laterally using legitimate administrative tools, typically via RDP.

Persistence
Employs techniques such as Kerberoasting and pass-the-hash, and tools like Mimikatz to extract credentials. Known to create new domain accounts to maintain access.

Command and Control
Utilizes remote access tools including AnyDesk, RustDesk, Ngrok, and Cloudflare Tunnel.

Exfiltration
Uses tools such as FileZilla, WinRAR, WinSCP, and Rclone. Data is exfiltrated via protocols like FTP and SFTP, or through cloud storage services such as Mega.

Darktrace’s Coverage of Akira ransomware

Reconnaissance

Darktrace first detected of unusual network activity around 05:10 UTC, when a desktop device was observed performing a network scan and making an unusual number of DCE-RPC requests to the endpoint mapper (epmapper) service. Network scans are typically used to identify open ports, while querying the epmapper service can reveal exposed RPC services on the network.

Multiple other devices were also later seen with similar reconnaissance activity, and use of the Advanced IP Scanner tool, indicated by connections to the domain advanced-ip-scanner[.]com.

Lateral movement

Shortly after the initial reconnaissance, the same desktop device exhibited unusual use of administrative tools. Darktrace observed the user agent “Ruby WinRM Client” and the URI “/wsman” as the device initiated a rare outbound Windows Remote Management (WinRM) connection to two domain controllers (REDACTED-dc1 and REDACTED-dc2). WinRM is a Microsoft service that uses the WS-Management (WSMan) protocol to enable remote management and control of network devices.

Darktrace also observed the desktop device connecting to an ESXi device (REDACTED-esxi1) via RDP using an LDAP service credential, likely with administrative privileges.

Credential access

At around 06:26 UTC, the desktop device was seen fetching an Active Directory certificate from the domain controller (REDACTED-dc1) by making a DCE-RPC request to the ICertPassage service. Shortly after, the device made a Kerberos login using the administrative credential.

Figure 3: Darktrace’s detection of the of anomalous certificate download and subsequent Kerberos login.

Further investigation into the device’s event logs revealed a chain of connections that Darktrace’s researchers believe demonstrates a credential access technique known as “UnPAC the hash.”

This method begins with pre-authentication using Kerberos’ Public Key Cryptography for Initial Authentication (PKINIT), allowing the client to use an X.509 certificate to obtain a Ticket Granting Ticket (TGT) from the Key Distribution Center (KDC) instead of a password.

The next stage involves User-to-User (U2U) authentication when requesting a Service Ticket (ST) from the KDC. Within Darktrace's visibility of this traffic, U2U was indicated by the client and service principal names within the ST request being identical. Because PKINIT was used earlier, the returned ST contains the NTLM hash of the credential, which can then be extracted and abused for lateral movement or privilege escalation [11].

Flowchart of Kerberos PKINIT pre-authentication and U2U authentication [12].
Figure 4: Flowchart of Kerberos PKINIT pre-authentication and U2U authentication [12]
Figure 5: Device event log showing the Kerberos Login and Kerberos Ticket events

Analysis of the desktop device’s event logs revealed a repeated sequence of suspicious activity across multiple credentials. Each sequence included a DCE-RPC ICertPassage request to download a certificate, followed by a Kerberos login event indicating PKINIT pre-authentication, and then a Kerberos ticket event consistent with User-to-User (U2U) authentication.

Darktrace identified this pattern as highly unusual. Cyber AI Analyst determined that the device used at least 15 different credentials for Kerberos logins over the course of the attack.

By compromising multiple credentials, the threat actor likely aimed to escalate privileges and facilitate further malicious activity, including lateral movement. One of the credentials obtained via the “UnPAC the hash” technique was later observed being used in an RDP session to the domain controller (REDACTED-dc2).

C2 / Additional tooling

At 06:44 UTC, the domain controller (REDACTED-dc2) was observed initiating a connection to temp[.]sh, a temporary cloud hosting service. Open-source intelligence (OSINT) reporting indicates that this service is commonly used by threat actors to host and distribute malicious payloads, including ransomware [13].

Shortly afterward, the ESXi device was observed downloading an executable named “vmwaretools” from the rare external endpoint 137.184.243[.]69, using the user agent “Wget.” The repeated outbound connections to this IP suggest potential command-and-control (C2) activity.

Cyber AI Analyst investigation into the suspicious file download and suspected C2 activity between the ESXI device and the external endpoint 137.184.243[.]69.
Figure 6: Cyber AI Analyst investigation into the suspicious file download and suspected C2 activity between the ESXI device and the external endpoint 137.184.243[.]69.
Packet capture (PCAP) of connections between the ESXi device and 137.184.243[.]69.
Figure 7: Packet capture (PCAP) of connections between the ESXi device and 137.184.243[.]69.

Data exfiltration

The first signs of data exfiltration were observed at around 7:00 UTC. Both the domain controller (REDACTED-dc2) and a likely SonicWall VPN device were seen uploading approximately 2 GB of data via SSH to the rare external endpoint 66.165.243[.]39 (AS29802 HVC-AS). OSINT sources have since identified this IP as an indicator of compromise (IoC) associated with the Akira ransomware group, known to use it for data exfiltration [14].

Cyber AI Analyst incident view highlighting multiple unusual events across several devices on August 20. Notably, it includes the “Unusual External Data Transfer” event, which corresponds to the anomalous 2 GB data upload to the known Akira-associated endpoint 66.165.243[.]39.
Figure 8: Cyber AI Analyst incident view highlighting multiple unusual events across several devices on August 20. Notably, it includes the “Unusual External Data Transfer” event, which corresponds to the anomalous 2 GB data upload to the known Akira-associated endpoint 66.165.243[.]39.

Cyber AI Analyst

Throughout the course of the attack, Darktrace’s Cyber AI Analyst autonomously investigated the anomalous activity as it unfolded and correlated related events into a single, cohesive incident. Rather than treating each alert as isolated, Cyber AI Analyst linked them together to reveal the broader narrative of compromise. This holistic view enabled the customer to understand the full scope of the attack, including all associated activities and affected assets that might otherwise have been dismissed as unrelated.

Overview of Cyber AI Analyst’s investigation, correlating all related internal and external security events across affected devices into a single pane of glass.
Figure 9: Overview of Cyber AI Analyst’s investigation, correlating all related internal and external security events across affected devices into a single pane of glass.

Containing the attack

In response to the multiple anomalous activities observed across the network, Darktrace's Autonomous Response initiated targeted mitigation actions to contain the attack. These included:

  • Blocking connections to known malicious or rare external endpoints, such as 137.184.243[.]69, 66.165.243[.]39, and advanced-ip-scanner[.]com.
  • Blocking internal traffic to sensitive ports, including 88 (Kerberos), 3389 (RDP), and 49339 (DCE-RPC), to disrupt lateral movement and credential abuse.
  • Enforcing a block on all outgoing connections from affected devices to contain potential data exfiltration and C2 activity.
Autonomous Response actions taken by Darktrace on an affected device, including the blocking of malicious external endpoints and internal service ports.
Figure 10: Autonomous Response actions taken by Darktrace on an affected device, including the blocking of malicious external endpoints and internal service ports.

Managed Detection and Response

As this customer was an MDR subscriber, multiple Enhanced Monitoring alerts—high-fidelity models designed to detect activity indicative of compromise—were triggered across the network. These alerts prompted immediate investigation by Darktrace’s SOC team.

Upon determining that the activity was likely linked to an Akira ransomware attack, Darktrace analysts swiftly acted to contain the threat. At around 08:05 UTC, devices suspected of being compromised were quarantined, and the customer was promptly notified, enabling them to begin their own remediation procedures without delay.

A wider campaign?

Darktrace’s SOC and Threat Research teams identified at least three additional incidents likely linked to the same campaign. All targeted organizations were based in the US, spanning various industries, and each have indications of using SonicWall VPN, indicating it had likely been targeted for initial access.

Across these incidents, similar patterns emerged. In each case, a suspicious executable named “vmwaretools” was downloaded from the endpoint 85.239.52[.]96 using the user agent “Wget”, bearing some resemblance to the file downloads seen in the incident described here. Data exfiltration was also observed via SSH to the endpoints 107.155.69[.]42 and 107.155.93[.]154, both of which belong to the same ASN also seen in the incident described in this blog: S29802 HVC-AS. Notably, 107.155.93[.]154 has been reported in OSINT as an indicator associated with Akira ransomware activity [15]. Further recent Akira ransomware cases have been observed involving SonicWall VPN, where no similar executable file downloads were observed, but SSH exfiltration to the same ASN was. These overlapping and non-overlapping TTPs may reflect the blurring lines between different affiliates operating under the same RaaS.

Lessons from the campaign

This campaign by Akira ransomware actors underscores the critical importance of maintaining up-to-date patching practices. Threat actors continue to exploit previously disclosed vulnerabilities, not just zero-days, highlighting the need for ongoing vigilance even after patches are released. It also demonstrates how misconfigurations and overlooked weaknesses can be leveraged for initial access or privilege escalation, even in otherwise well-maintained environments.

Darktrace’s observations further reveal that ransomware actors are increasingly relying on legitimate administrative tools, such as WinRM, to blend in with normal network activity and evade detection. In addition to previously documented Kerberos-based credential access techniques like Kerberoasting and pass-the-hash, this campaign featured the use of UnPAC the hash to extract NTLM hashes via PKINIT and U2U authentication for lateral movement or privilege escalation.

Credit to Emily Megan Lim (Senior Cyber Analyst), Vivek Rajan (Senior Cyber Analyst), Ryan Traill (Analyst Content Lead), and Sam Lister (Specialist Security Researcher)

Appendices

Darktrace Model Detections

Anomalous Connection / Active Remote Desktop Tunnel

Anomalous Connection / Data Sent to Rare Domain

Anomalous Connection / New User Agent to IP Without Hostname

Anomalous Connection / Possible Data Staging and External Upload

Anomalous Connection / Rare WinRM Incoming

Anomalous Connection / Rare WinRM Outgoing

Anomalous Connection / Uncommon 1 GiB Outbound

Anomalous Connection / Unusual Admin RDP Session

Anomalous Connection / Unusual Incoming Long Remote Desktop Session

Anomalous Connection / Unusual Incoming Long SSH Session

Anomalous Connection / Unusual Long SSH Session

Anomalous File / EXE from Rare External Location

Anomalous Server Activity / Anomalous External Activity from Critical Network Device

Anomalous Server Activity / Outgoing from Server

Anomalous Server Activity / Rare External from Server

Compliance / Default Credential Usage

Compliance / High Priority Compliance Model Alert

Compliance / Outgoing NTLM Request from DC

Compliance / SSH to Rare External Destination

Compromise / Large Number of Suspicious Successful Connections

Compromise / Sustained TCP Beaconing Activity To Rare Endpoint

Device / Anomalous Certificate Download Activity

Device / Anomalous SSH Followed By Multiple Model Alerts

Device / Anonymous NTLM Logins

Device / Attack and Recon Tools

Device / ICMP Address Scan

Device / Large Number of Model Alerts

Device / Network Range Scan

Device / Network Scan

Device / New User Agent To Internal Server

Device / Possible SMB/NTLM Brute Force

Device / Possible SMB/NTLM Reconnaissance

Device / RDP Scan

Device / Reverse DNS Sweep

Device / Suspicious SMB Scanning Activity

Device / UDP Enumeration

Unusual Activity / Unusual External Data to New Endpoint

Unusual Activity / Unusual External Data Transfer

User / Multiple Uncommon New Credentials on Device

User / New Admin Credentials on Client

User / New Admin Credentials on Server

Enhanced Monitoring Models

Compromise / Anomalous Certificate Download and Kerberos Login

Device / Initial Attack Chain Activity

Device / Large Number of Model Alerts from Critical Network Device

Device / Multiple Lateral Movement Model Alerts

Device / Suspicious Network Scan Activity

Unusual Activity / Enhanced Unusual External Data Transfer

Antigena/Autonomous Response Models

Antigena / Network / External Threat / Antigena File then New Outbound Block

Antigena / Network / External Threat / Antigena Suspicious Activity Block

Antigena / Network / External Threat / Antigena Suspicious File Block

Antigena / Network / Insider Threat / Antigena Large Data Volume Outbound Block

Antigena / Network / Insider Threat / Antigena Network Scan Block

Antigena / Network / Insider Threat / Antigena Unusual Privileged User Activities Block

Antigena / Network / Manual / Quarantine Device

Antigena / Network / Significant Anomaly / Antigena Alerts Over Time Block

Antigena / Network / Significant Anomaly / Antigena Controlled and Model Alert

Antigena / Network / Significant Anomaly / Antigena Enhanced Monitoring from Client Block

Antigena / Network / Significant Anomaly / Antigena Enhanced Monitoring from Server Block

Antigena / Network / Significant Anomaly / Antigena Significant Anomaly from Client Block

Antigena / Network / Significant Anomaly / Antigena Significant Server Anomaly Block

Antigena / Network / Significant Anomaly / Repeated Antigena Alerts

List of Indicators of Compromise (IoCs)

·      66.165.243[.]39 – IP Address – Data exfiltration endpoint

·      107.155.69[.]42 – IP Address – Probable data exfiltration endpoint

·      107.155.93[.]154 – IP Address – Likely Data exfiltration endpoint

·      137.184.126[.]86 – IP Address – Possible C2 endpoint

·      85.239.52[.]96 – IP Address – Likely C2 endpoint

·      hxxp://85.239.52[.]96:8000/vmwarecli  – URL – File download

·      hxxp://137.184.126[.]86:8080/vmwaretools – URL – File download

MITRE ATT&CK Mapping

Initial Access – T1190 – Exploit Public-Facing Application

Reconnaissance – T1590.002 – Gather Victim Network Information: DNS

Reconnaissance – T1590.005 – Gather Victim Network Information: IP Addresses

Reconnaissance – T1592.004 – Gather Victim Host Information: Client Configurations

Reconnaissance – T1595 – Active Scanning

Discovery – T1018 – Remote System Discovery

Discovery – T1046 – Network Service Discovery

Discovery – T1083 – File and Directory Discovery

Discovery – T1135 – Network Share Discovery

Lateral Movement – T1021.001 – Remote Services: Remote Desktop Protocol

Lateral Movement – T1021.004 – Remote Services: SSH

Lateral Movement – T1021.006 – Remote Services: Windows Remote Management

Lateral Movement – T1550.002 – Use Alternate Authentication Material: Pass the Hash

Lateral Movement – T1550.003 – Use Alternate Authentication Material: Pass the Ticket

Credential Access – T1110.001 – Brute Force: Password Guessing

Credential Access – T1649 – Steal or Forge Authentication Certificates

Persistence, Privilege Escalation – T1078 – Valid Accounts

Resource Development – T1588.001 – Obtain Capabilities: Malware

Command and Control – T1071.001 – Application Layer Protocol: Web Protocols

Command and Control – T1105 – Ingress Tool Transfer

Command and Control – T1573 – Encrypted Channel

Collection – T1074 – Data Staged

Exfiltration – T1041 – Exfiltration Over C2 Channel

Exfiltration – T1048 – Exfiltration Over Alternative Protocol

References

[1] https://thehackernews.com/2025/08/sonicwall-investigating-potential-ssl.html

[2] https://www.sonicwall.com/support/notices/gen-7-and-newer-sonicwall-firewalls-sslvpn-recent-threat-activity/250804095336430

[3] https://psirt.global.sonicwall.com/vuln-detail/SNWLID-2024-0015

[4] https://arcticwolf.com/resources/blog/arctic-wolf-observes-akira-ransomware-campaign-targeting-sonicwall-sslvpn-accounts/

[5] https://www.rapid7.com/blog/post/dr-akira-ransomware-group-utilizing-sonicwall-devices-for-initial-access/

[6] https://www.ic3.gov/AnnualReport/Reports/2024_IC3Report.pdf

[7] https://www.cisa.gov/news-events/cybersecurity-advisories/aa24-109a

[8] https://blog.talosintelligence.com/akira-ransomware-continues-to-evolve/

[9] https://www.ransomware.live/map?year=2025&q=akira

[10] https://attack.mitre.org/groups/G1024/
[11] https://labs.lares.com/fear-kerberos-pt2/#UNPAC

[12] https://www.thehacker.recipes/ad/movement/kerberos/unpac-the-hash

[13] https://www.s-rminform.com/latest-thinking/derailing-akira-cyber-threat-intelligence)

[14] https://fieldeffect.com/blog/update-akira-ransomware-group-targets-sonicwall-vpn-appliances

[15] https://arcticwolf.com/resources/blog/arctic-wolf-observes-july-2025-uptick-in-akira-ransomware-activity-targeting-sonicwall-ssl-vpn/

Get the latest insights on emerging cyber threats

This report explores the latest trends shaping the cybersecurity landscape and what defenders need to know in 2026.

Written by
Emily Megan Lim
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
Emily Megan Lim
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