Human-in-the-Loop (HITL)
Security pattern requiring a human reviewer to approve sensitive operations before an agent executes them. Single-use approval consumed on the next matching call.
Human-in-the-Loop (HITL) is a security pattern requiring a human reviewer to explicitly approve sensitive operations before an AI agent executes them. The approval is single-use — consumed on the next matching call — preventing agents from accumulating standing permission.
When to use HITL
- Destructive operations (delete, revoke, refund)
- High-value transactions (wire transfers, contract signing)
- Access to sensitive data categories (PII, financial records)
- Operations outside normal business hours or patterns
HITL vs consent
Consent is a one-time authorization for a tool category and trust tier. HITL is a per-operation approval for individual sensitive calls. Consent says "I trust this agent to use GitHub." HITL says "show me every time it tries to push to main."
Related terms
Security mechanism that scans tool descriptions, arguments, and results for prompt injection payloads before AI agents process them. Detects instruction-override patterns and hidden unicode.
Graduated levels of trust assigned to AI agents: low trust requires explicit consent for every tool, high trust auto-approves common patterns. Capped by the user's RBAC permissions.
User-facing interface where a human explicitly authorizes an AI agent to access specific tools within defined boundaries. Time-boxed, revocable, and access-aware.