4 ms·
> (...) and the only communication channel would be towards me (enforced with things like API key permissions). > This should prevent any kind of leaks due to
by e12e 8mo ago
> (...) and the only communication channel would be towards me (enforced with things like API key permissions).
> This should prevent any kind of leaks due to prompt injection, right ?
It might be harder than you think. Any conditional fetch of an URL or DNS query could reveal some information.
- LelouBil 8mo agoDNS Queries are fine, and also conditional URL fetches, as long as they are not arbitrary, should be okay too. I don't mind the agent searching my GMail using keywords from some discord private messages for example, but I would mind if it did a web search because it could give anything to the search result URLs.