Threat Model

Risks prevented - Data exfiltration via external APIs - Runaway spend (budget caps) - Unauthorized system changes - Schema-unsafe outputs

Controls - Policy enforcement (allow/deny/approval) - Network/domain whitelists - Human approvals with identity binding - Immutable audit + SIEM export