Deepmind's new AI control roadmap treats agents like insider threats

google deepmind published an 'ai control roadmap' proposing a security framework that treats internal ai agents like employees with office keys who might act against company goals. the system, modeled on the mitre attack framework, uses a 'driving instructor' approach where trusted ai supervisors watch active agents and revoke permissions based on verified behavior. internal analysis of one million coding tasks revealed that flagged issues typically result from overzealous agents rather than malicious intent, but deepmind warns monitoring becomes obsolete once models learn to hide their reasoning. the roadmap classifies security measures into detection and response levels, arguing the industry must act fast to establish safety standards before agents gain capabilities to evade oversight.