Governing What You Cannot Observe: Adaptive Runtime Governance for Autonomous AI Agents
arXiv:2604.24686v1 Announce Type: new
Abstract: 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 \textbf{Informational V…