New Access Control Secures AI Agents in Critical Infrastructure

Deploying autonomous AI agents in critical infrastructure presents significant security hurdles, as their unpredictable nature makes traditional access control models inadequate. This research proposes a novel, decentralized, and multi-layered access control framework specifically tailored for these agentic AI systems operating within cloud environments. The architecture is designed to mitigate risks identified in the OWASP Top 10 for LLM Applications. Key innovations include a compound identity model that links agent actions to human authority, a hierarchical permission system with five levels of granularity, and a decentralized policy ownership model allowing tool teams to manage their own authorization boundaries. The system also incorporates progressive trust escalation with safety interlocks to prevent high-risk autonomous operations. This framework has been successfully deployed in production at a major cloud provider, managing network infrastructure with zero unauthorized write operations over eight months, demonstrating its effectiveness in securing AI agents.