AgentWall: Runtime Safety Layer for Local AI Agents

AgentWall is a runtime safety and observability layer designed for local AI agents, introduced in a paper on arXiv (2605.16265v1). It intercepts every proposed agent action before execution, evaluating it against an explicit declarative policy. This addresses the gap where existing AI safety work focuses on model alignment and input filtering but not on real-time action control. The system is critical for local environments where agents interact with filesystems, credentials, and infrastructure. AgentWall ensures that unsafe or adversarially manipulated behaviors are blocked at runtime, providing a new layer of protection for autonomous agents.