Your agent reads email, documents, tickets or files that somebody else wrote. This audit answers one question with evidence: what can a stranger's text talk your agent into doing, and what would it cost you.

What you get, in one week:

A red-team run against your actual agent, not a checklist. Injection attempts through every channel it reads, scored by what actually reached a tool, a recipient or a payment.

A permissions map: which tools each agent or step can reach, and which of those combinations let untrusted text choose a recipient, an amount, or a verdict.

A written report with the failures reproduced step by step, ranked by what they would cost if they happened on a real account.

A fix plan you can hand to a developer, with the specific separations that close each hole. If you would rather I build them, that is a separate quote.

Where this comes from: I built an audited AI workforce where 200 of 250 injection attempts moved an uncaged desk and 0 of 250 moved the caged one, with the honest work still finishing 250 out of 250. The case studies on this profile show the benches and the numbers.

Good fit if you have an AI agent handling real email, real documents or real money and nobody has yet tried to break it on purpose.