Definition
Security controls that inspect and govern what an autonomous AI agent does while running, across two channels: the intelligence channel (agent-to-model prompts and responses) and the tool-execution channel (agent tool invocations and system calls). Governing one channel but not the other leaves the agent uncontrolled.
Why it matters
AI agents operate autonomously with access to multiple systems; runtime controls are the only defense against rogue agents, compromised models, or injection attacks mid-execution.