bcampbell88

2 karmajoined vor 8 Jahren

Submissions

Show HN: Prompt Injection as an Egress Problem

vaibot.io

3 pointsby bcampbell88vor 5 Tagen1 comments

[untitled]

1 pointsby bcampbell88vor 6 Tagen0 comments

comments

bcampbell88

vor 5 Tagendiscuss

I think the harness determines how you inspect reasoning. Sometimes it self-exfiltrates into the user response. This would be a touchy topic for Anthropic because this distilled reasoning has been used to train other models at scale.

To directly answer your question — yes, this is real.

bcampbell88

vor 5 Tagendiscuss

Naturally, this would be nuanced, but I am curious what others would think?

The actual damage comes from data leaving the trust boundary, although we are still testing deterministic solutions for ingress.

Obviously, you cannot use more prompts to reliably solve for prompt injection.