What changed
Google DeepMind published an AI Control roadmap that treats capable internal agents as potential insider threats and layers monitoring, prevention and response on top of conventional security and model alignment. The framework maps controls to an agent's ability to evade detection and cause harm, and reports an internal asynchronous monitor informed by one million coding-agent trajectories.
Why it matters
Agent governance gains a capability-scaled operating model: delayed transcript review may suit reversible work, while high-consequence actions require synchronous prevention and measurable monitoring coverage, recall and response time.
What to do
Threat-model one deployed agent as an insider, classify its reversible and irreversible actions, measure monitoring coverage and alert recall, and move high-impact actions behind synchronous blocking and an independently controlled recovery path.
Caveat
The roadmap and million-trajectory analysis come from Google and describe internal systems. The public article does not provide an independent production evaluation, full dataset or proof that supervisors remain reliable against stronger or oversight-aware agents.