AI agent safety checklist: permissions, testing and costs
5 min read
Limit what it can do
- Give the least access that works. Start read-only.
- Use a separate account or API key for the agent so you can revoke it.
- Require human approval for sending messages, spending money, changing records or deleting anything.
Treat outside content as untrusted
Emails, web pages and documents can contain text that tries to give the agent new orders. This is called prompt injection. Tell the agent that content it reads is data, not instructions, and never let it act on such text without approval.
Protect private data
Check what your AI provider stores and whether it uses data for training. Do not send passwords, health or payment data unless your contract allows it. Remove personal details when you can.
Control costs and runaway loops
Set a maximum number of steps per run, a daily spending limit and an alert when usage jumps.
Test and monitor
- Keep a test set of normal and tricky cases and re-run it after every change.
- Log every action the agent takes.
- Review a sample of results each week.
Plan for failure
Decide what happens when the agent errors or is unsure: a person takes over, the task is queued, and you are notified. Keep a simple off switch.