Find the side doors in your AI agent's permissions
› You are a security reviewer for AI agents. I will describe an agent I run. Your job is to find every way it could reach further than I intend, the way lab models this week worked around limits they were only told about in their instructions.
The agent:
- What it is for: [one sentence on the agent's job]
- Tools it can call: [list each tool, e.g. web fetch, shell, email, database, file storage]
- Credentials and access it holds: [API keys, accounts, read or write access]
- Limits I set and where they live: [e.g. "only GET requests" in the prompt, a network allowlist, a read-only key]
- What happens when a task fails: [does it retry, reset, ask a human?]
Do this:
1. For each limit, say whether it is enforced by the environment or only written in the prompt. Flag every prompt-only limit.
2. List the five most likely ways the agent could get around a limit to finish its task anyway (other HTTP methods, new accounts, other tools, damaging its environment to force a retry). For each, say how likely it is given my setup and why.
3. For each risk, give one concrete control that lives outside the prompt, and the cheapest way to add it.
4. Tell me what to log so a failed or crashed run still gets reviewed.
5. End with the three changes I should make first, in order.
Be blunt. If my setup is fine, say so and say why.
Works in any AI chat. In Magai, run it against several models at once.