AI Control Roadmap (securing internal systems)
AI Control Roadmap is Google DeepMind's defense-in-depth plan for running capable agents internally even if their alignment is imperfect.
Treats internal agents as potential insider threats. Alignment stays the primary defense, with system-level controls and permissions that grow with verified behavior. Pitched as a template for the industry.
- Date
- Thursday 18 June 2026
- Lab
- Google DeepMind
- Kind
- paper
- Access
- paper only
A policy and framework post. The page is dated June 18, 2026, while the site's RSS feed lists it under June 16.
Sources
This record was checked and corrected against its sources on 6 October 2026. How we check