AI智能体安全审计模板

Instrument AI agent security decisions with Telemetry. Use /skill.md and this Telemetry API key: YOUR_API_KEY At the authorization boundary, log agent_tool_authorization_decided with: event_id, run_id, tool_call_id, workflow, agent_name, tool_name, action_class, risk_level, decision, reason_code, policy_version, release, environment, and timestamp_utc. Use controlled values for action_class, risk_level, decision, and reason_code. Emit the authorization decision before execution, then log agent_tool_completed separately with the same correlation identifiers, status, duration_ms, and a bounded error_type. Create reviewed SQL for decision volume, denial rate, approval-required volume, policy-version changes, and allowed actions that later fail. Keep absolute counts beside rates. Do not log prompts, completions, credentials, authorization headers, tool arguments, tool results, retrieved content, customer data, or free-form policy explanations. Keep enforcement 中 the policy layer. Use Telemetry to query the resulting events.