AI Safety & Trust and Safety
Autonomous Agent Policy
Purpose
Governs design, deployment, and oversight of autonomous agents, including WSAI CODE.
Scope
All agentic systems taking multi-step actions with limited human intervention.
Definitions
Autonomous Agent.
Rigorous Definition of "Consequential Action" A Consequential Action is any agent-initiated action that: (a) modifies data or systems outside a disposable sandbox (e.g., production deployments, database writes, repository merges to protected branches); (b) incurs financial cost or executes a transaction; (c) sends external communications (emails, API calls to third-party services, public posts); (d) modifies access permissions or credentials; or (e) cannot be trivially and fully reversed within 5 minutes. Any action meeting one or more criteria requires a Consequential Action Checkpoint.
Policy Statements
Technical Enforcement Mandate Consequential Action Checkpoints must be implemented as a formal API-level approval flow: the agent cannot proceed past the checkpoint without an explicit, logged user or designated-approver confirmation returned through the platform API — checkpoints cannot be satisfied by the agent's own self-assessment. Each agent operates under a formal permission schema (a machine-readable manifest declaring exactly which actions, systems, and data scopes it may access) reviewed and approved by Security before deployment. All actions are logged in cryptographically signed, hash-chained audit logs (see Security Policy Doc. 11 and Prompt Injection Policy Doc. 19). Long-running agent loops are subject to resource/time limits and mandatory periodic human-on-the-loop check-ins (minimum every 30 minutes of continuous execution or 20 consequential-adjacent steps, whichever is sooner). Agentic features undergo mandatory penetration testing (see Doc. 19) before release and quarterly thereafter.
Agents must operate within sandboxed execution environments (e.g., WSAI CODE's dedicated cloud compute) and support emergency interruption at any point.
Responsibilities
Research Engineering owns agent architecture and permission schemas; Security approves permission manifests pre-deployment; AI Safety Research Lead reviews new agent capabilities.
Compliance Requirements
Agentic releases pass Autonomous Agent Risk Assessments under the AI Safety Framework; permission schemas re-certified quarterly.
Enforcement
Agents exhibiting unsafe behavior or checkpoint bypass attempts are suspended pending investigation; repeated bypass attempts trigger automatic circuit-breaker shutdown.
Reporting Process
Exceptions Process
Expanded autonomy permissions require Governance Review Board approval and a documented, time-limited pilot period with enhanced monitoring.
Review Process
Quarterly.