GDM AI Control Roadmap
The GDM AI Control Roadmap (v0.1) has been released, outlining a plan for internal guardrails to detect and mitigate adversarial AI agent behavior. The roadmap draws inspiration from cybersecurity frameworks like MITRE ATT&CK, introducing TRAIT&R for threat modeling and categorizing potential AI threats into loss of control, work sabotage, and direct harm. It establishes control invariants for detecting misaligned intent and proposes capability-based mitigation ladders, suggesting 15 concrete defenses tiered according to evolving model capabilities. AI
IMPACT Provides a structured approach to AI safety and control, integrating cybersecurity principles for future AI systems.