Audit your AI agent for an emergency brake
› Act as a blunt reviewer of AI agent safety and cost. I will describe one AI agent or automation I run. Audit it against these five checks and give each a PASS, FAIL or UNKNOWN with one line of reasoning:
1. Separation: can the agent change its own instructions, permissions, tools or logs? It should not be able to.
2. Outside controls: which limits (network, files, spending, accounts) are enforced outside the model, and which rely on the model behaving?
3. Evidence: does every action that touches the outside world (sending, submitting, buying, writing to a system) leave a record a person can read later and the agent cannot edit?
4. Brake: can a named person pause or stop a task mid-step without redeploying? Who, and how?
5. Cheap decisions: list every step where the agent is really making a yes/no or pick-one call (routing, triage, pass/fail checks). Flag which could go to a smaller, cheaper model with a confidence threshold, and which need a human below that threshold.
Then give me the three fixes that cut the most risk for the least work, in order, each with a first step I can do this week. Ask me questions first if my description leaves a check UNKNOWN.
My agent: [describe what it does, which model it uses, what tools and accounts it can reach, and who watches it]
Works in any AI chat. In Magai, run it against several models at once.