Securing Agents and Autonomous Behavior

This CIS Companion Guide addresses the expanded security surface when AI systems move from text prediction to autonomous action, focusing on risks at the agent layer (code/tool execution, record updates, workflow orchestration), memory and RAG-induced behavioral risks, multi-agent orchestration hazards, and the importance of deployment/runtime ownership; it prescribes controls such as strict tool allowlisting, capability scoping, identity and access controls, human-in-the-loop for high-impact actions, memory protection and validation, and deterministic guardrails.