Google DeepMind guides its AI agents using the AI Control Roadmap

Google DeepMind launches the AI Control Roadmap to monitor Google AI agents like Gemini Spark, treating unaligned models as insider threats via trusted AIs.

Google DeepMind is formalizing how it monitors AI agents deployed within Google, in a framework dubbed the AI Control Roadmap. The principle reverses the usual assumption: rather than presuming an agent is aligned with our intentions, it starts from the premise that it might not be and could act unexpectedly. Model alignment, posited as the first line of defense, is then supplemented by a system-level security layer, designed to hold even when alignment fails.

Unreliable agents are treated as insider threats, similar to an employee already granted access, following a modeling approach based on the MITRE ATT&CK standard. Other trusted AIs act as supervisors: they continuously review the agent's reasoning, actions, and plans, and block a harmful maneuver before it can cause damage.

DeepMind, however, points out a limitation: if models learn to conceal their reasoning, reading their chain of thought will no longer suffice, and it will be necessary to examine their behavior, or even their internal workings. To calibrate these monitors, the team says it has sifted through a million code agent trajectories, enough to equip the Gemini Spark agent with real-time monitoring, particularly against unintentional data deletion. Most incidents, it clarifies, stem less from malicious intent than from overzealousness or misinterpretation. A framework for decision-makers, Three Layers of Agent Security, accompanies the publication.