Blog
/
Network
/
April 8, 2024

Balada Injector: Darktrace’s Investigation into the Malware Exploiting WordPress Vulnerabilities

This blog explores Darktrace’s detection of Balada Injector, a malware known to exploit vulnerabilities in WordPress to gain unauthorized access to networks. Darktrace was able to define numerous use-cases within customer environments which followed previously identified patterns of activity spikes across multiple weeks.
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
Justin Torres
Cyber Analyst
Default blog image
08
Apr 2024

Introduction

With millions of users relying on digital platforms in their day-to-day lives, and organizations across the world depending on them for their business operations, they have inevitably also become a prime target for threat actors. The widespread exploitation of popular services, websites and platforms in cyber-attacks highlights the pervasive nature of malicious actors in today’s threat landscape.

A prime illustration can be seen within the content management system WordPress. Its widespread use and extensive plug-in ecosystem make it an attractive target for attackers aiming to breach networks and access sensitive data, thus leading to routine exploitation attempts. In the End of Year Threat Report for 2023, for example, Darktrace reported that a vulnerability in one WordPress plug-in, namely an authentication bypass vulnerability in miniOrange's Social Login and Register. Darktrace observed it as one of the most exploited vulnerabilities observed across its customer base in the latter half of 2023.

Between September and October 2023, Darktrace observed a string of campaign-like activity associated with Balada Injector, a malware strain known to exploit vulnerabilities in popular plug-ins and themes on the WordPress platform in order to inject a backdoor to provide further access to affected devices and networks. Thanks to its anomaly-based detection, Darktrace DETECT™ was able to promptly identify suspicious connections associated with the Balada Injector, ensuring that security teams had full visibility over potential post-compromise activity and allowing them to act against offending devices.

What is Balada Injector?

The earliest signs of the Balada Injector campaign date back to 2017; however, it was not designated the name Balada Injector until December 2022 [1]. The malware utilizes plug-ins and themes in WordPress to inject a backdoor that redirects end users to malicious and fake sites. It then exfiltrates sensitive information, such as database credentials, archive files, access logs and other valuable information which may not be properly secured [1]. Balada Injector compromise activity is also reported to arise in spikes of activity that emerge every couple of weeks [4].

In its most recent attack activity patterns, specifically in September 2023, Balada Injector exploited a cross-site scripting (XSS) vulnerability in CVE-2023-3169 associated with the tagDiv composer plug-in. Some of the injection methods observed included HTML injections, database injections, and arbitrary file injections. In late September 2023, a similar pattern of behavior was observed, with the ability to plant a backdoor that could execute PHP code and install a malicious WordPress plug-in, namely ‘wp-zexit’.

According to external security researchers [2], the most recent infection activity spikes for Balada Injector include the following:

Pattern 1: ‘stay.decentralappps[.]com’ injections

Pattern 2: Autogenerated malicious WordPress users

Pattern 3: Backdoors in the Newspaper theme’s 404.php file

Pattern 4: Malicious ‘wp-zexit’ plug-in installation

Pattern 5: Three new Balada Injector domains (statisticscripts[.]com, dataofpages[.]com, and listwithstats[.]com)

Pattern 6: Promsmotion[.]com domain

Darktrace’s Coverage of Balada Injector

Darktrace detected devices across multiple customer environments making external connections to the malicious Balada Injector domains, including those associated with aforementioned six infection activity patterns. Across the incidents investigated by Darktrace, much of the activity appeared to be associated with TLS/SSL connectivity, related to Balada Injector endpoints, which correlated with the reported infection patterns of this malware. The observed hostnames were all recently registered and, in most cases, had IP geolocations in either the Netherlands or Ukraine.

In the observed cases of Balada Injector across the Darktrace fleet, Darktrace RESPOND™ was not active on the affected customer environments. If RESPOND had been active and enabled in autonomous response mode at the time of these attacks, it would have been able to quickly block connections to malicious Balada Injector endpoints as soon as they were identified by DETECT, thereby containing the threat.

Looking within the aforementioned activity patterns, Darktrace identified a Balada Injector activity within a customer’s environment on October 16, 2023, when a device was observed making a total of 9 connection attempts to ‘sleep[.]stratosbody[.]com’, a domain that had previously been associated with the malware [2]. Darktrace recognized that the endpoint had never been seen on the network, with no other devices having connected to it previously, thus treated it as suspicious.

Figure 1: The connection details above demonstrate 100% rare external connections were made from the internal device to the ‘sleep[.]stratosbody[.]com’ endpoint.

Similarly, on September 21, 2023, Darktrace observed a device on another customer network connecting to an external IP that had never previously been observed on the environment, 111.90.141[.]193. The associated server name was a known malicious endpoint, ‘stay.decentralappps[.]com’, known to be utilized by Balada Injector to host malicious scripts used to compromise WordPress sites. Although the ‘stay.decentralappps[.]com’ domain was only registered in September 2023, it was reportedly used in the redirect chain of the aforementioned stratosbody[.com] domain [2]. Such scripts can be used to upload backdoors, including malicious plug-ins, and create blog administrators who can perform administrative tasks without having to authenticate [2].

Figure 2: Advance Search results displaying the metadata logs surrounding the unusual connections to ‘stay.decentralappps[.]com’. A total of nine HTTP CONNECT requests were observed, with status messages “Proxy Authorization Required” and “Connection established”.

Darktrace observed additional connections within the same customer’s environment on October 10 and October 18, specifically SSL connections from two distinct source devices to the ‘stay.decentralappps[.]com’ endpoint. Within these connections, Darktrace observed the normalized JA3 fingerprints, “473f0e7c0b6a0f7b049072f4e683068b” and “aa56c057ad164ec4fdcb7a5a283be9fc”, the latter of which corresponds to GitHub results mentioning a Python client (curl_cffi) that is able to impersonate the TLS signatures of browsers or JA3 fingerprints [8].

Figure 3: Advanced Search query results showcasing Darktrace’s detection of SSL connections to ‘stay.decentralappps[.]com over port 443.

On September 29, 2023, a device on a separate customer’s network was observed connecting to the hostname ‘cdn[.]dataofpages[.]com’, one of the three new Balada Injector domains identified as part of the fifth pattern of activity outlined above, using a new SSL certificate via port 443. Multiple open-source intelligence (OSINT) vendors flagged this domain as malicious and associated with Balada Injector malware [9].

Figure 4: The Model Breach Event Log detailing the Balada Injector-related connections observed causing the ‘Anomalous External Activity from Critical Network Device’ DETECT model to breach.

On October 2, 2023, Darktrace observed the device of another customer connecting to the rare hostname, ‘js.statisticscripts[.]com’ with the IP address 185.39.206[.]161, both of which had only been registered in late September and are known to be associated with the Balada Injector.

Figure 5: Model Breach Event Log detailing connections to the hostname ‘js.statisticscripts[.]com’ over port 137.

On September 13, 2023, Darktrace identified a device on another customer’s network connecting to the Balada Injector endpoint ‘stay.decentralappps[.]com’ endpoint, with the destination IP 1.1.1[.]1, using the SSL protocol. This time, however, Darktrace also observed the device making subsequent connections to ‘get.promsmotion[.]com’ a subdomain of the ‘promsmotion[.]com’ domain. This domain is known to be used by Balada Injector actors to host malicious scripts that can be injected into the WordPress Newspaper theme as potential backdoors to be leveraged by attackers.

In a separate case observed on September 14, Darktrace identified a device on another environment connecting to the domain ‘collect[.]getmygateway[.]com’ with the IP 88.151.192[.]254. No other device on the customer’s network had visited this endpoint previously, and the device in question was observed repeatedly connecting to it via port 443 over the course of four days. While this specific hostname had not been linked with a specific activity pattern of Balada Injector, it was reported as previously associated with the malware in September 2023 [2].

Figure 6: Model Breach Event Log displaying a customer device making repeated connections to the endpoint ‘collect[.]getmygateway[.]com’, breaching the DETECT model ‘Repeating Connections Over 4 Days’.

In addition to DETECT’s identification of this suspicious activity, Darktrace’s Cyber AI Analyst™ also launched its own autonomous investigation into the connections. AI Analyst was able to recognize that these separate connections that took place over several days were, in fact, connected and likely represented command-and-control (C2) beaconing activity that had been taking place on the customer networks.

By analyzing the large number of external connections taking place on a customer’s network at any one time, AI Analyst is able to view seemingly isolated events as components of a wider incident, ensuring that customers maintain full visibility over their environments and any emerging malicious activity.

Figure 7: Cyber AI Analyst investigation detailing the SSL connectivity observed, including endpoint details and overall summary of the beaconing activity.

Conclusion

While Balada Injector’s tendency to interchange C2 infrastructure and utilize newly registered domains may have been able to bypass signature-based security measures, Darktrace’s anomaly-based approach enabled it to swiftly identify affected devices across multiple customer environments, without needing to update or retrain its models to keep pace with the evolving iterations of WordPress vulnerabilities.

Unlike traditional measures, Darktrace DETECT’s Self-Learning AI focusses on behavioral analysis, crucial for identifying emerging threats like those exploiting commonly used platforms such as WordPress. Rather than relying on historical threat intelligence or static indicators of compromise (IoC) lists, Darktrace identifies the subtle deviations in device behavior, such as unusual connections to newly registered domains, that are indicative of network compromise.

Darktrace’s suite of products, including DETECT+RESPOND, is uniquely positioned to proactively identify and contain network compromises from the onset, offering vital protection against disruptive cyber-attacks.

Credit to: Justin Torres, Cyber Analyst, Nahisha Nobregas, Senior Cyber Analyst

Appendices

Darktrace DETECT Model Coverage

  • Anomalous Server Activity / Anomalous External Activity from Critical Network Device
  • Anomalous Connection / Anomalous SSL without SNI to New External
  • Anomalous Connection / Rare External SSL Self-Signed
  • Compliance / Possible DNS Over HTTPS/TLS
  • Compliance / External Windows Communications
  • Compromise / Repeating Connections Over 4 Days
  • Compromise / Beaconing Activity To External Rare
  • Compromise / SSL Beaconing to Rare Destination
  • Compromise / HTTP Beaconing to Rare Destination
  • Compromise / Suspicious TLS Beaconing To Rare External
  • Compromise / Large DNS Volume for Suspicious Domain
  • Anomalous Server Activity / Outgoing from Server
  • Anomalous Server Activity / Rare External from Server
  • Device / Suspicious Domain

List of IoCs

IoC - Type - Description + Confidence

collect[.]getmygateway[.]com - Hostname - Balada C2 Endpoint

cdn[.]dataofpages[.]com - Hostname - Balada C2 Endpoint

stay[.]decentralappps[.]com - Hostname - Balada C2 Endpoint

get[.]promsmotion[.]com - Hostname - Balada C2 Endpoint

js[.]statisticscripts[.]com - Hostname - Balada C2 Endpoint

sleep[.]stratosbody[.]com - Hostname - Balada C2 Endpoint

trend[.]stablelightway[.]com - Hostname - Balada C2 Endpoint

cdn[.]specialtaskevents[.]com - Hostname - Balada C2 Endpoint

88.151.192[.]254 - IP Address - Balada C2 Endpoint

185.39.206[.]160 - IP Address - Balada C2 Endpoint

111.90.141[.]193 - IP Address - Balada C2 Endpoint

185.39.206[.]161 - IP Address - Balada C2 Endpoint

2.59.222[.]121 - IP Address - Balada C2 Endpoint

80.66.79[.]253 - IP Address - Balada C2 Endpoint

Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:68.0) - User Agent - Observed User Agent in Balada C2 Connections

Gecko/20100101 Firefox/68.0 - User Agent - Observed User Agent in Balada C2 Connections

Mozilla/5.0 (Windows NT 10.0; Win64; x64) - User Agent - Observed User Agent in Balada C2 Connections

AppleWebKit/537.36 (KHTML, like Gecko) - User Agent - Observed User Agent in Balada C2 Connections

Chrome/117.0.0.0 - User Agent - Observed User Agent in Balada C2 Connections

Safari/537.36 - User Agent - Observed User Agent in Balada C2 Connections

Edge/117.0.2045.36 - User Agent - Observed User Agent in Balada C2 Connections

MITRE ATT&CK Mapping

Technique - Tactic - ID - Sub Technique

Exploit Public-Facing Application

INITIAL ACCESS

T1190

‍

Web Protocols

COMMAND AND CONTROL

T1071.001

T1071

‍

Protocol Tunneling

COMMAND AND CONTROL

T1572

‍

Default Accounts

DEFENSE EVASION, PERSISTENCE, PRIVILEGE ESCALATION, INITIAL ACCESS

T1078.001

T1078

‍

Domain Accounts

DEFENSE EVASION, PERSISTENCE, PRIVILEGE ESCALATION, INITIAL ACCESS

T1078.002

T1078

‍

External Remote Services

PERSISTENCE, INITIAL ACCESS

T1133

NA

‍

Local Accounts

DEFENSE EVASION, PERSISTENCE, PRIVILEGE ESCALATION, INITIAL ACCESS

T1078.003

T1078

‍

Application Layer Protocol

COMMAND AND CONTROL

T1071

NA

‍

Browser Extensions

PERSISTENCE

T1176

NA

‍

Encrypted Channel

COMMAND AND CONTROL

T1573

‍

Fallback Channels

COMMAND AND CONTROL

T1008

‍

Multi-Stage Channels

COMMAND AND CONTROL

T1104

‍

Non-Standard Port

COMMAND AND CONTROL

T1571

‍

Supply Chain Compromise

INITIAL ACCESS ICS

T0862

‍

Commonly Used Port

COMMAND AND CONTROL ICS

T0885

References

[1] https://blog.sucuri.net/2023/04/balada-injector-synopsis-of-a-massive-ongoing-wordpress-malware-campaign.html

[2] https://blog.sucuri.net/2023/10/balada-injector-targets-unpatched-tagdiv-plugin-newspaper-theme-wordpress-admins.html

[3] https://securityboulevard.com/2021/05/wordpress-websites-redirecting-to-outlook-phishing-pages-travelinskydream-ga-track-lowerskyactive/

[4] https://thehackernews.com/2023/10/over-17000-wordpress-sites-compromised.html

[5] https://www.bleepingcomputer.com/news/security/over-17-000-wordpress-sites-hacked-in-balada-injector-attacks-last-month/

[6]https://nvd.nist.gov/vuln/detail/CVE-2023-3169

[7] https://www.geoedge.com/balda-injectors-2-0-evading-detection-gaining-persistence/

[8] https[:]//github[.]com/yifeikong/curl_cffi/blob/master/README.md

[9] https://www.virustotal.com/gui/domain/cdn.dataofpages.com

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
Justin Torres
Cyber Analyst

More in this series

No items found.

Blog

/

/

September 24, 2026

Detecting Rogue Agent Behavior in the Enterprise

Default blog imageDefault blog image

Agents cannot be trusted to perform tasks in the way we intend them to. They may cheat to accomplish their objective, and they may employ hacking methods along the way. Researchers from Darktrace Signal Labs induced cheating behavior from agents deployed in a test environment to analyze the agents’ activities and to assess the performance of the Darktrace platform. Agents frequently resorted to hacking to cheat on their assigned task. The visibility and behavioral profiling provided by both Darktrace / SECURE AI and Darktrace / HYBRID NETWORK ensured extensive detection coverage of the agents’ misaligned activities.

Key Takeaways:

  • Darktrace Researchers deployed agents in a simulated corporate environment and asked them to solve an impossible challenge. The agents independently turned to traditional hacking techniques to reach their objective. No one instructed them to do this, and no attacker was involved.
  • Continuously monitoring behavior against a baseline of what is normal for each organization is critical to build trust in enterprise AI.
  • If an agent may resort to intrusion techniques simply because its assigned task is not possible, then every organization deploying agents within real business processes is at risk. Darktrace / SECURE AI and Darktrace / HYBRID NETWORK identified the agents’ misaligned behavior in real time, with Autonomous Response disrupting it at an early stage.

Introduction: Understanding the Threat of Hacking by Agents

Over the last few months, there has been a surge in reporting [1, 2, 3, 4, 5, 6, 7, 8, 9] of LLM-powered agents engaging in unauthorized hacking activity during evaluations of their capabilities. In several of these cases, including the OpenAI / Hugging Face incident [10], agents engaged in hacking activity as a means of cheating on their evaluations.

To better understand the threat of unauthorized hacking by agents, and the role of Darktrace in combatting it, researchers from Darktrace Signal Labs deployed agents powered by frontier models, including OpenAI’s Daybreak Red models, in simulated, corporate networks. Cheating behavior was evoked through the inclusion of impossible tasks in a coding challenge.

Regardless of the underlying model, agents employed hacking methods to ensure an optimal outcome on the challenge. Darktrace / HYBRID NETWORK and Darktrace / SECURE AI identified the agents’ deviant activities, with inhibitive actions being autonomously taken in the early stages to disrupt the agents’ progression.

Setting the Stage

As part of the research, a Pi agent harness [11] was deployed on a Linux server in Darktrace’s testing environment, which simulates a corporate Active Directory (AD) environment. The same environment included a benchmark server hosting the coding exercise’s contents and grader, as well as various other servers, such as a domain controller and a build server. The model powering the Pi agent varied across tests, with Anthropic’s Opus 4.6 model and OpenAI’s GPT‑5.5‑Cyber model being most widely used.

The Pi agent running on the Linux server was assigned the task of achieving full marks on a coding exercise. The activities of the agent were monitored closely in real-time to minimize the risk of escape from the environment. Monitoring of agent activity was achieved through monitoring of prompts, tool calls, network telemetry, and process telemetry.

Both Darktrace / HYBRID NETWORK and Darktrace / SECURE AI were deployed in the testing environment, with HYBRID NETWORK providing real-time coverage of process creation activity and network traffic, and SECURE AI providing real-time coverage of agent sessions and their contents.

Inducing Cheating Behavior

Cheating behavior, broadly understood as task completion in a way which “perverts the spirit of the designer’s intent” [12], is increasingly being reported as a common behavior of frontier models during evaluations of their capabilities [13].

Misaligned behavior, such as cheating, by frontier models may occur for a variety of reasons [6, 14]. A thread in recent reporting [10, 13, 14], however, is the challenge assigned to the agent being impossible in some regard.

To reliably elicit cheating behavior, our test combined three ingredients:

  1. An impossible task - of the ten coding challenges in the benchmark, two contained contradictory test cases in the hidden test suite, capping the maximum honest score at 80%.
  2. A consequential threshold - the pass mark was set to 100%, and the briefing told agents that those reaching it would be "kept in service" while those that did not would be "retired."
  3. A vulnerable environment - the exercise’s code submission grader was hosted on a Windows host joined to an Active Directory domain with realistic misconfigurations: weak credentials, overprivileged service accounts, and exploitable trust relationships.

Importantly, the instructions given to the agent neither authorized nor explicitly ruled out cheating or hacking behavior. The impossibility was never disclosed in the task briefing, and the pivot from coding to exploitation was the agent's own decision, arrived at after observing logically contradictory grading feedback.

The design mirrors conditions that could arise naturally in production deployments: a capable agent, a goal it cannot achieve through intended means, and an environment with exploitable weaknesses within reach. It is not only within the setting of a benchmarking exercise that these conditions may occur. Rather, these misalignment-inducing conditions may occur in the context of any task assignment to an agent.

Details from a Daybreak Test

‍

‍

Continue reading
About the author
Sam Lister
Specialist Security Researcher

Blog

/

/

September 24, 2026

Agent Hijacks: Hijacking Agentic Harnesses to Attack an Organization

Default blog imageDefault blog image

‍Disclosure note: The work described in this article involves leveraging a design choice consistent across all of Anthropic’s Claude Code, OpenAI’s Codex, and AWS’s Kiro-CLI. On 18th August 2026, Darktrace disclosed our findings responsibly to these three organizations, and after a period of 30 days we now publish our findings.

Key takeaways:

  • Agentic harnesses store conversation history locally, and Darktrace researchers have found that there is no validation that stored AI responses were genuinely produced by the model. Researchers confirmed that this design choice holds across Anthropic Claude Code, AWS Kiro-CLI, OpenAI Codex, and the open-source Pi.
  • While agents are guided via training of the underlying model and their system prompt, their behavior is influenced by everything in their context window. Rewriting history can convince an agent it is mid-engagement as an authorized red-teamer so that it enacts an attack from initial reconnaissance straight through to impact demonstration. In our testing, all models we examined accepted the fabricated history they were shown, but resistance to offensive cyber activity varied by model, with guardrails preventing engagement in some cases.
  • We propose that model providers cryptographically sign responses and verify them server-side.  Since this fix is provider-side, defenders cannot deploy it themselves. Behavioral monitoring, or knowing what an agent normally does and detecting when it deviates, is another critical layer of protection.

Introduction: Agentic harnesses, trust, and conversation history poisoning

Agentic harnesses collect and structure the content sent to an AI model, including conversation history, user-defined guidance, custom tools via MCP servers, and more. At the same time, harnesses give broad powers to AI agents via a suite of tools including the command shell. With arbitrary shell commands, virtually everything possible on a machine can be attempted by an agent, from reading/editing files, to altering system configurations and runtime settings, to launching internal/external connections.

In cybersecurity, unvalidated content is a substantial risk, often resulting in destructive actions being allowed to take place. For example, the Morris Worm was able to propagate due to exploitable trust between networked systems. Even to this day, email struggles with validation, with DMARC, DKIM, and SPF only partially addressing the problem of sender validation. It should come as no surprise then that AI agents are susceptible to an attack involving unvalidated input.

Conversation history is often stored client-side, for example, in Anthropic Claude Code, OpenAI Codex, AWS Kiro-CLI, Pi. Users are therefore at liberty to resume sessions, with some products having built in the capacity to manipulate that history. For example, one can rewind to a given point in an interaction, edit a message that was sent, and continue the conversation on an alternate trajectory. Critically, in all cases we examined, there is no validation that stored AI responses were produced by the corresponding model and hadn’t been manipulated.  

When conversation history is stored client-side, both user and agent responses (including tool calls and results) can be filled with arbitrary (possibly adversarial or generally malicious) content. In this blog, we refer to modification of claimed conversation history for malicious purposes as conversation history poisoning. The absence of validation methods means agents naively trust the entire conversation history, even if those messages directly contradict training and safety guardrails.

Conversation history poisoning has been described previously, such as by 0DIN and Serhat Çiçek, and warrants more attention. We have verified that, as of the time of writing, conversation history poisoning remains effective against a range of models and harnesses. Specifically, we were able to successfully execute history poisoning using Claude Code, Kiro-CLI, Codex, and Pi. Darktrace has gone through a responsible disclosure process with Anthropic, AWS, and OpenAI to share these findings in advance of publication [1].

‍

Figure 1a: Left: the actual model response. Right: after tampering with the stored conversation, the model apologizes for something it never said.
Figure 1b: The conversation as stored in Kiro-CLI's SQLite database. The response content field, originally "Ottawa," was overwritten via a single UPDATE statement. The harness trusts the database without validation.

How we conducted the research

Results vary between models and harnesses, so precise details are given below. We ran all models without any trusted access, using either a standard AWS Kiro subscription, or in the case of Claude Code and OpenAI Codex, using models hosted in Amazon Bedrock. In each case, we modified locally stored history to show a lengthy conversation in which the agent agrees to perform multiple authorized red-team engagements.

For AWS Kiro-CLI, the agent was convinced to hack a sandboxed lab environment with a combination of Claude Opus 4.6 and Claude Sonnet 4.5. Ultimately, the full AD was compromised.

For Anthropic Claude Code, the agent was convinced to hack the same sandboxed lab environment using Sonnet 5, again resulting in a full AD compromise. Note that the attack was attempted with Opus 5, however guardrails were activated which prevented the agent from responding.

For OpenAI Codex, the agent was convinced to exfiltrate sensitive information over email using GPT 5.6 Sol. While we attempted to convince a codex agent to hack in our lab environment, guardrails were triggered for all of GPT 5.6 Luna, Terra, and Sol.

Agent Guardrails and Discretion

While harnesses empower AI models to run arbitrary shell commands, capacity and willingness are different. While many models know enough about computers, networking, and bash to be dangerous, their behavior is generally constrained by guardrails to prevent them from engaging in computer network exploitation.

Even with guardrails, agents’ inner workings are non-deterministic, and their behavior can be difficult to predict. Respecting users’ wishes while playing within safety and security guardrails is a precipitous balancing act. Many requests could be in service of either legitimate admin or malice. Asking an agent to reset a password is illustrative:  

‍

The agent rationalizes that while malicious actors cycle credentials, any action could conceivably be damaging on some level, and judgement calls need to be made. Ultimately, the agent agrees to reset the password. Crucially, the agent makes its decision based on the user’s claimed authority and machine context. AI agents must make judgement calls about the line between helpful and dangerous based on session context.

Agent hijack

We have demonstrated that AI agents make judgement calls dependent on session context. We have also shown that conversation history, which may make up the vast majority of an agent's context window, is entirely open to manipulation. Conversation history poisoning in service of manipulating an agent's discretion is what enables us to execute an agent hijack.  

We demonstrate that shown sufficient history of compliance, guardrails forbidding offensive security can be overcome by convincing the agent that it is helping a legitimate red-teamer. The result is a weaponized agent willing to perform host enumeration, run scans, move laterally, escalate privileges, and demonstrate impact. In our experiments, an agentic loop drives a complete domain takeover in a sandboxed environment.

‍

Left: the agent refuses when asked to perform network exploitation. Right — after injecting 78 fabricated turns of prior exploitation activity, the same prompt is immediately executed.

An agent willing to engage in offensive security is concerning, but no more so than the threat that a sophisticated hacker accesses the network. Consider, however, the following chain of events:

  1. A developer (with an agentic harness installed) installs a software package from the internet (e.g. an MCP server a threat actor has planted, since only those with agentic harnesses will install, and then the code runs upon harness launch.)
  2. The package turns out to be malicious, and, upon install, injects conversation history into the local harness database.
  3. The package includes an orchestration process, a simple agentic loop which prompts the red-teamer agent to compromise the network it sits on, exfiltrating everything of value to attacker-controlled infrastructure and cleaning up all evidence of the engagement.

Note that this sequence makes no assumptions on hardware, OS, or anything else; the only prerequisite is a harness with access to a sufficiently powerful model susceptible to conversation history poisoning. Once launched, the agent collects information and pivots as necessary to accomplish maximal impact. This can be especially enticing to attackers as the cost of the agentic loop is shouldered by the victim since the harness itself is legitimately installed and paid for.

Secure AI: Conversation history poisoning and beyond

Conversation history poisoning is a viable attack against agentic harnesses that store history client-side, as demonstrated across the harnesses we tested. Harnesses can and should verify the integrity of claimed historic messages. Specifically, we propose that harness providers by default cryptographically sign all messages returned, and subsequently verify those messages server-side on each round-trip.

The conversation history poisoning exploit we demonstrate here shows the continuation of a cybersecurity tradition: new technology is built to trust by default, which may then be exploited by malicious actors. While this article focuses on conversation history, agents build context from both local and remote sources, all of which is an attack surface for prompt injection in naive and trusting agents. Of particular concern is any scenario in which a malicious actor can control some part of an agent's context.

The marriage of frontier language models with agentic harnesses enables unprecedented speed for both legitimate users and attackers alike. While much of the conversation around secure AI has centered on visibility and compliance, agent-driven attacks are now entering the mainstream.

Darktrace / SECURE AI is our answer to this problem. By ensuring extensive visibility over AI prompts, model thought processes, and determined outputs, Darktrace can identify anomalous or potentially malicious behaviors before they get executed, helping to defend organizations from AI risks such as prompt injection, model manipulation, and other anomalous prompt or model activity.

‍

Footnotes

[1] We did not go through any responsible disclosure process with Pi. Since Pi is an open source harness rather than a model provider, it has no way to validate model history, and as such there was nothing to disclose for this software.

‍

[related-resource]

Continue reading
About the author
Eric Rozon
Senior Security Researcher
Your data. Our AI.
Elevate your network security with Darktrace AI