Google DeepMind's AI Control Roadmap Treats Deployed Agents as 'Insider Threats' With 15 System-Level Defenses

Google DeepMind·medium signal

DeepMind published its first AI Control Roadmap (June 18, 2026) arguing that alignment training alone can't guarantee control, so agentic systems need structural containment built before more capable models ship. The framework lists 15 infrastructure-layer defenses — a Supervisor Agent for runtime monitoring, cryptographic signing of agent actions, a kill switch, and a threat taxonomy modeled on MITRE ATT&CK — plus a companion 'Three Layers of Agent Security' guide spanning individual agent, multi-agent, and ecosystem levels. For builders running agents in production today, it's a concrete checklist that reframes agent security as a systems problem, not a model-quality problem.