AI Agent Risk Checklist
A pre-deployment checklist for AI agents that can call tools or take actions, focused on permissions, observability, reversibility and escalation.
Define the job
- Write the goal and stopping condition in plain language.
- List the tools and data the agent genuinely needs.
- Prefer deterministic automation where the task is already deterministic.
Restrict authority
- Apply least-privilege permissions.
- Require confirmation for irreversible, financial or external actions.
- Separate test and production credentials.
Observe behaviour
- Log tool calls, actions and important intermediate decisions.
- Test malformed inputs, unavailable tools and conflicting instructions.
- Measure failure recovery, not just successful demo runs.
Plan recovery
- Make actions reversible where practical.
- Define a human escalation route and kill switch.
- Re-test after model, prompt, tool or permission changes.
Evidence behind this checklist
- Department for Science, Innovation and Technology · AI Adoption Research · 2026-02-13 · Primary source
- UK Government · Code of Practice for the Cyber Security of AI · 2025-01-31 · Primary source