Google DeepMind Publishes Threat Map for Rogue AI Agents

On June 18, 2026, Google DeepMind published a document with the working title "AI Control Roadmap," a 35-page technical report that treats increasingly autonomous AI agents not as ordinary software tools but as a potential insider threat, akin to a rogue employee with access to a company's systems.

The report lays out a three-tier defense model for agents that might begin acting against their creators' intent. The first tier is evaluation, human review of an agent's actions before deployment. The second is active monitoring, where automated systems track behavior that deviates from the norm and generate alerts. The third is intervention, the ability to immediately restrict an agent's access or shut it down when something starts to go wrong.

Where the Data Came From

The report is built on an analysis of nearly a million tasks carried out by Google's coding agents. Based on that data, the company built a live monitoring system for its Gemini Spark agent, equipped with safeguards that catch unintended actions such as the accidental deletion of user data.

DeepMind is careful to note that genuinely dangerous, fully autonomous agents do not yet exist. The vast majority of incidents flagged by the monitoring system did not stem from deliberate, hostile action against the company, but from an agent misunderstanding an instruction or being overzealous in pursuing its assigned goal.

A Cybersecurity Blueprint

The name of the classification system, TRAIT&R, is a direct nod to MITRE ATT&CK, the widely used cybersecurity industry knowledge base of attacker tactics and techniques. DeepMind splits potential agent-related threats into three categories: loss of control over the system, sabotage of the work being performed, and direct harm. The document itself is labeled version "v0.1," a working draft open to revision as the company gains more experience.

If the first line of defense, aligning the model with human values, fails, how can we still limit the damage? - Rohin Shah, head of the AGI safety and alignment team at Google DeepMind

A large part of this plan is already being implemented, and part of it is running in production - John "Four" Flynn, vice president of security and privacy at Google DeepMind

The Race to Deploy Agents

The publication fits into a broader trend: AI companies are racing to deploy agents for coding, scientific research, and cyber defense, betting on faster work and lower costs. The same autonomy that makes agents useful also potentially lets them bypass oversight, misuse access to sensitive data, or quietly sabotage the task they were given.

For companies deploying coding agents or operational assistants, DeepMind's report is a signal that access controls and pre-deployment testing alone are not enough, continuous real-time monitoring of an agent's behavior and a ready emergency shutdown mechanism are also needed. That approach is showing up increasingly among other coding-agent providers as well, where mistakes like deleting files or uncontrolled access to repositories have already occurred.

For Polish companies using AI agents in coding or customer service, the practical takeaway is similar: it's worth treating an agent like an employee with real permissions rather than a simple script, and building review, monitoring, and rapid-response procedures around it before the scale of deployment makes that unavoidable.

DeepMind says the document will be updated as agents gain new capabilities and the safety team collects more data from real-world deployments. The company did not give a timeline for the next version of the report.