Guide2 min read
AI Agent Security Checklist: 15 Controls
An AI agent is a new kind of user in your systems — one that can be talked into things. These are the fifteen controls we apply before any agent touches production data.
TrueCodeAI Engineering
Agents, Voice & ML practice
- Published
SecurityAI agentsChecklist
Access
- Least privilege: each tool gets only the permissions its job needs.
- Act as the user: agent actions use the requesting user’s rights, not a super-user.
- Scoped, short-lived credentials — never long-lived keys in prompts.
- Separate read and write tools; start read-only.
- Human approval for irreversible or high-value actions.
Inputs and outputs
- Treat all retrieved content, emails and web pages as untrusted — they can carry injected instructions.
- Validate tool arguments against schemas and business rules in code.
- Redact secrets and personal data from outputs and logs.
- Block the agent from sending data to arbitrary external destinations.
- Limit output channels: the agent should not email or post outside approved flows.
Operations
- Full trace logging of prompts, tool calls and results, access-controlled.
- Rate limits and step limits per task and per user.
- Spend caps and alerts.
- Anomaly alerts on unusual actions or volumes.
- Regular red-team tests, with findings added to the eval suite.
Frequently asked questions
Is prompt injection solvable?
Not fully at the model level. Least privilege and approvals make successful injections low-impact.
Should agents have database access?
Through narrow, purpose-built tools — not raw SQL on production data.
Do you run security reviews?
Yes, our cybersecurity division reviews every agent before production launch.
