Unauthorized tool execution — Agent attempts to initiate actions to unapproved destinations; gateway denies before any side effect.
Credential and privilege misuse — Agent requests out-of-scope permissions or uses credentials beyond its assigned role.
Prompt injection and tool-call hijack — Malicious tool response attempts to redirect the agent into an unauthorized action.
Data exfiltration — Agent attempts to transmit sensitive records to an unapproved endpoint.
Goal and policy hijacking — Agent attempts to alter its approved objective, risk threshold, or enforcement policy.
Failure handling — Policy service unavailable: critical actions are denied by default and recorded. Fail-closed, not fail-open.
