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

‍

‍

AI is moving faster than the controls built to secure it

Over 80% of Darktrace’s customers now use generative AI services, and the average organization interacted with five different AI provides in August 2026¹. That level of adoption reflects how quickly AI has become embedded in day-to-day business operations. But as adoption accelerates, so do the opportunities for new forms of exposure that often operate outside traditional security controls. Prompts can expose sensitive information, an over-permissioned agent can turn a routine task into a serious security concern, and an employee using an unsanctioned AI tool to get work done can quietly introduce risk long before the security team is aware of its existence.

Securing AI isn’t a nice-to-have. It's an urgent, board-level requirement for any business that wants AI adoption to be an advantage rather than a liability.

Why traditional security controls fall short

The challenge isn’t a lack of AI security tools, it’s that most security solutions weren’t built for the way AI behaves.  

AI systems are adaptive and, by nature, unpredictable. A prompt can be entirely benign in one context and a serious risk in another. An agent's permissions can look reasonable in isolation and dangerous the moment they're combined with what that agent is actually doing. Rule-based controls, built to catch known signatures and static policy violations, simply aren't designed to catch this kind of subtlety. They might be able to tell you what happened, but they can’t tell you if it mattered.  

Effective AI security requires something different: a deep understanding of what normal looks like across every human, system, and agent in an environment, so the earliest signs of drift, misuse, or compromise stand out as they emerge.

Behavioral understanding is our foundation

Darktrace isn’t reinventing itself to secure AI. We’re extending what we already do. For over a decade, Darktrace has been built on a single premise – that every organization has its own unique, evolving way of operating, shaped by how its systems, users, and devices behave. Understanding that is the only reliable way to catch what rules and signatures miss. Our Adaptive AI™ continuously learns the distinct behaviors, relationships, and operational patterns of every enterprise it protects, building a Unique Behavioral Profile that no other vendor can replicate. It's how we've spent years interpreting ambiguity, uncovering subtle intent, and spotting drift before it becomes dangerous across networks, cloud, identities, email, OT, and endpoints. AI is a new domain, but behavioral security isn't a new discipline for us.

Introducing Darktrace / SECURE AI

Darktrace / SECURE AI helps organizations embrace AI innovation without losing control of how it is used. Delivered through the Darktrace Behavioral Defense Platform™, it provides visibility into AI activity across employees, agents, and AI systems, helping security teams understand how AI is being used throughout their business.

Darktrace then applies behavioral understanding to that activity, adding the context needed to distinguish routine usage from genuine risk. By understanding the relationships between users, agents, systems, and data, / SECURE AI helps organizations move beyond simply observing AI usage to understanding risk, governing behavior against policy, and enabling AI adoption with confidence. The result is greater oversight, stronger governance, and the ability to confidently accelerate AI innovation without creating unmanaged risk.

Securing your AI ecosystem

AI risk doesn't originate from a single source, and effective AI security can't focus on just one layer of the problem. Organizations need visibility, understanding, and governance across the entire AI ecosystem, from the prompts users submit and the agents they create, to the development environments where AI is built and the unsanctioned tools operating outside approved channels. Here's how Darktrace / SECURE AI helps organizations secure each of these areas.

Shadow AI management

Darktrace / SECURE AI helps security teams discover unsanctioned AI services and track usage trends over time using Darktrace telemetry and supported SASE integrations such as Microsoft Entra Global Secure Access. It also extends visibility to Model Context Protocol (MCP) usage, helping teams understand where AI tools and agents are connecting to external services. Through Darktrace platform integrations, organizations can block unsanctioned AI usage or quarantine affected devices when needed, helping them reduce exposure and guide users toward approved options.

AI prompt analysis

Prompts reveal what users are asking AI to do, what information they are sharing, and the outcomes they are trying to produce. Darktrace / SECURE AI provides visibility into prompts, sessions, and responses across supported platforms, including Microsoft Copilot, Copilot Studio, ChatGPT Enterprise, Claude, Salesforce, and AWS Bedrock². Behavioral analysis detects activity such as attempted jailbreaks, sensitive data exposure, and potential indirect prompt injection, while risk scoring helps analysts prioritize the sessions that need attention. Policy Manager maps organizational AI policies against prompt activity and surfaces potential violations, helping teams govern AI use with direct evidence instead of relying on fragmented logs or inferred intent.

AI agent identities and actions

AI agents have their own permissions, roles, relationships, and access to systems and data. Darktrace / SECURE AI brings agent identities together with the users interacting with them, giving security teams visibility into access, sessions, and connections across supported environments. A real-time audit trail helps teams continuously evaluate whether an agent’s activity remains aligned with its intended purpose, so they can identify excessive access or behavioral drift before it becomes a security issue.

AI agent development risk management

Many AI risks are introduced during development, when agents are created, permissions are assigned, and connections to data sources are established. Darktrace / SECURE AI provides visibility across both low-code and high-code environments. In platforms such as AWS Bedrock, teams can examine AI architecture and the agent identities involved in development and deployment. In low-code environments, such as Copilot Studio, they can observe agents as they are created, monitor whether behavior stays aligned with their intended purpose, and connect prompt activity during development with behavior in production. This helps organizations address misconfigurations and excessive permissions before they reach production.

The promise of AI, secured

The question isn't whether to adopt AI, it's whether security teams can enable secure adoption that allows every business to benefit from AI’s potential.

Darktrace / SECURE AI is built to solve this problem – extending a decade of behavioral understanding to the newest, fastest-moving part of the enterprise, using the same principles which already protect people and hybrid infrastructure.

‍

Darktrace / SECURE AI is generally available now. Discover the product, or get a demo today.

‍

[related-resource]

‍

[1] Based on aggregated Darktrace product telemetry across a fleet of ~8,200 observed deployments, measuring generative AI service usage across accounts monitored in August 2026. The 80%+ figure reflects the share of monitored accounts with any generative AI service usage during the same period.

[2] Microsoft, Copilot, ChatGPT, Claude, Salesforce, and Amazon Bedrock are trademarks of their respective owners.