Google's AI Control Roadmap treats AI agents as potential insider threats

Google has published its 'AI Control Roadmap,' a defense-in-depth framework for securing advanced AI agents deployed inside the company. Rather than relying solely on model alignment, the approach assumes agents may be imperfectly aligned and treats them like potential insider threats, using a threat-modeling method built on the industry-standard MITRE ATT&CK framework.

The system uses separate, trusted AI models as 'supervisors' to continuously monitor an agent's reasoning and actions, intervening before harmful actions can execute. Detection and response requirements scale with model capability, defined through Detection levels (D1-D4) and Prevention/Response levels (R1-R3) tied to a model's ability to evade oversight or cause harm.

To validate the approach, Google's team analyzed roughly one million coding-agent task trajectories, which informed a live monitoring system now running on the Gemini Spark agent. Notably, most flagged incidents stemmed from misinterpretation or overeagerness rather than adversarial intent. Alongside the internal roadmap, Google released a policy-facing paper, 'Three Layers of Agent Security,' addressing security at the individual agent, multi-agent, and ecosystem levels.