Google DeepMind has published its AI Control Roadmap, a defense-in-depth security framework for managing advanced AI agents deployed internally. The approach treats AI agents as potential insider threats — even if alignment is imperfect — and layers traditional cybersecurity practices (sandboxing, prompt injection resistance) with AI-powered supervisors that monitor agent reasoning and actions in real time. The roadmap maps security protocols to measurable AI capability milestones, scaling from asynchronous transcript review for low-risk actions to synchronous real-time blocking for high-risk ones. DeepMind has already analyzed one million coding agent trajectories to refine behavioral detection, moving beyond keyword filtering. A companion policy paper, 'Three Layers of Agent Security,' outlines how the industry, policymakers, and academia should collaborate on agent security standards.
