Governing What You Cannot Observe: Adaptive Runtime Governance for Autonomous AI Agents

German Marin, Jatin Chaudhary

Autonomous AI agents can remain fully authorized and still become unsafe as behavior drifts, adversaries adapt, and decision patterns shift without any code change. We propose the…