Untrusted content isolation

Prompt injection fires at read time, not at some later "use" step. Opening the file is the moment of exposure.

Why: once a tool-capable context has read attacker-controlled text, the compromise has already happened; guarding the later action is too late.

How to apply: route untrusted content through a sandboxed reader that has no tools and no ability to act, and let only its distilled summary cross back into the acting context.