Google DeepMind Publishes 'AI Control Roadmap' to Defend Against Rogue AI Agents

Treating Your Own AI as an Insider Threat

Google DeepMind has published what it calls an "AI Control Roadmap," a 35-page technical framework that does something unusual: it treats the company's own AI agents not as trusted software tools, but as potential insider threats capable of sabotage, data exfiltration, or evading human oversight. Released on Thursday and labelled version 0.1, the document marks DeepMind's first formal attempt to build defences against AI systems that might one day work against the people running them.

Fifteen Layers of Defence

At the heart of the roadmap is a "defence-in-depth" approach made up of 15 layered safeguards. These range from simple network activity logs all the way up to real-time behavioural monitoring and automated access revocation. The core idea is to use trusted AI models as "supervisors" that continuously review a working agent's reasoning, actions, and plans, stepping in to block harmful behaviour before any damage is done.

DeepMind acknowledges a thorny problem: an agent's stated reasoning does not always reflect what the model is actually doing. To counter potential deception, the company proposes monitoring the activation patterns inside neural networks, comparing the technique to using an fMRI scan to read a person's thoughts in real time. DeepMind says it has already built an internal prototype that asynchronously monitors its own coding agents and escalates suspicious findings for human review.

A Cybersecurity Playbook for AI

Central to the framework is TRAIT&R, a Taxonomy of Rogue AI Tactics and Routines, modelled on MITRE ATT&CK, the widely used knowledge base of adversarial behaviours in cybersecurity. The taxonomy sorts threats into three classes: "loss of control," such as a model setting up unmonitored deployments; "work sabotage," such as agents degrading safety research; and "direct harm," such as exfiltrating model weights.

The roadmap scales its defences to capability. It calibrates detection levels based on how well a model can evade monitoring, and prevention levels based on how capable a model is of executing an attack. Safeguards progress from simply assessing what actions agents take, to monitoring and alerting on concerning behaviour, to infrastructure that can restrict access or disable an agent in real time.

Setting an Industry Standard

DeepMind released the roadmap publicly alongside a companion guide for policymakers, with the stated aim of establishing an industry standard. The company plans to fold the framework into its broader Frontier Safety Framework as it matures. Notably, DeepMind was careful to stress that genuinely hazardous autonomous agents do not yet exist, but argued the defences must be built before they arrive. The move lands amid a wider industry push on agent safety, with OpenAI, Ping Identity, and Akamai all recently unveiling their own methods to predict, authenticate, and monitor increasingly autonomous AI systems.