DeepMind's AI Control Roadmap treats deployed agents as insider threats.

Published June 18, the roadmap argues alignment training alone can't guarantee control, so agentic systems need structural containment built before more capable models ship. It 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. (Google DeepMind) For anyone running agents in production, it's a checklist that reframes agent security as a systems problem, not a model-quality one. That framing is correct and most teams are still treating it as a prompt problem.