Skip to content

Comment on Interactively Audit System Security with AIparent

Comments

swaxOP

The LLM only communicates with the shell so other users shouldn’t be able to leverage it. Maybe if the LLM reads a file that a lower privilege user has access to that’d be a route for injection into the context at least. There’s also a ‘command protection’ system we’re trying out that uses a separate secondary LLM to approve the commands of the primary.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.