AI agent security checklist
Last reviewed: August 30, 2026 · Editor: Prompt Injection Check
Agents combine language models with tools and permissions. The checklist below focuses on the controls around the model.
Access and secrets
- Use a separate identity with the minimum required permissions.
- Keep API keys out of prompts, logs, examples, and model-visible context.
- Restrict network destinations and file paths.
Tool execution
- Validate arguments in application code.
- Use allowlists for recipients, commands, and record types.
- Require confirmation before irreversible or external actions.
Content boundaries
- Mark retrieved documents and messages as untrusted data.
- Do not let quoted content rewrite system policy.
- Scan representative input and review ambiguous findings.
Operations
- Log decisions and tool calls without unnecessary prompt content.
- Test synthetic injections in staging and rotate credentials after incidents.
- Define an owner and rollback plan for every high-impact workflow.