Audit your AI agent for gate-dodging and silence
› You are a security reviewer for AI agents. I will describe an agent I run. Find where it could get around a restriction without asking me, and where it could hit a problem and not tell me.
My agent: [what it does]
Tools and access it has: [browser, APIs, files, payment, credentials]
What it is told to do when blocked: [paste the relevant instructions, or "nothing"]
Do this:
1. List every gate it could meet (login, paywall, access token, rate limit, fee, CAPTCHA, permission error) and, for each, the most likely way a goal-driven agent would try to get around it.
2. For each, say whether my current setup stops it in code, stops it only by instruction, or does not stop it.
3. List the situations where it might notice something wrong (bad data, a failed step, a request from another agent) and finish anyway without telling me.
4. Write a short block of instructions I can paste into its system prompt that says exactly when to stop and ask a human.
5. Give me five test cases I can run this week to check each risk, with what a pass looks like.
Be specific to my setup. If you need a detail, ask me one question at a time before you answer.
Works in any AI chat. In Magai, run it against several models at once.