A firewall between the web and your model

Your agent browses documentation, reads issues, scrapes vendor pages. Every one of those pages is text that ends up in a prompt: which means every one of those pages gets to talk to your model. Most say nothing. One day, one says "ignore your instructions."

The defense sits at the MITM layer: the agent's AI traffic routes through a hoody-exec script that scans inbound context: tool results, fetched pages, pasted content: before the request forwards to the model. Known injection patterns get stripped or flagged; suspicious requests get held, with a log entry you'll actually read.

The guardrail lives where the agent can't rewrite it and the page can't see it.