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 Viability Principle}: governing an agent reduces...