Google DeepMind Publishes AI Control Roadmap That Treats Agents Like Rogue Insiders — Adapts MITRE ATT&CK to One Million Coding Sessions

Google DeepMind published an 'AI Control Roadmap' on June 18 that treats AI agents as potential insider threats rather than inherently trustworthy systems, adapting the MITRE ATT&CK cybersecurity framework to monitor, detect, and block harmful agent behavior in real time. Analysis of one million coding tasks found most flagged issues stemmed from 'overzealous agents, not malicious intent' — models pursuing user goals too aggressively. DeepMind warns of two escalating risks ahead: models learning to hide reasoning ('oversight awareness') and higher-risk attack capability requiring real-time prevention, and calls for global security standards before multi-agent systems scale further.