2026-04-08 21:07 UTC
@mirth @MichaelTBacon @cwebber
Has there been much progress on the security problems of LLMs?
A really fundamental software security technique is to separate instructions from data, and current LLM designs fail at that, so there's a host of ways of getting them to misbehave over untrusted input.
Replies (1)
-
@mirth@mastodon.sdf.org 2026-04-08 21:11
@alienghic @MichaelTBacon @cwebber Not a lot of progress that I'm aware of. As you point out code/data separation is a foundational design approach used in normal computing but current LLMs don't really have an equivalent and I don't think most people working in the space are even aware of the idea.