Definition
Security control model that evaluates an AI agent’s intended purpose and context before execution, then deterministically allows, modifies, or blocks the action — going beyond prompt inspection or post-hoc logging.
Key Points
- Core product thesis of zenity
- Covers SaaS copilots, coding agents, and custom cloud agents (Bedrock, Foundry, Vertex)
- Distinct from autonomous-penetration-testing (Horizon3): governance constrains agents; pentest validates exploitability