RE: infosec.exchange/@mttaggart/11…
This is a really important read for understanding both what's under the hood with LLMs, but also why it is impossible to secure any #LLM.
In the thread, I posted a couple of extra links to experts explaining that last sentence better than I can. But read the article Taggart linked first.
#AI #TulipMania #infosec
This is a masterful demonstration of the unsecurable nature of LLMs, and how prompt injection by dedicated humans who know how to write will always win.Prompt Injection as Role Confusion
LLMs can't tell who's speaking. We show they identify roles by writing style, not tags, and exploit this with CoT Forgery, injecting fake reasoning that models mistake for their own thoughts.role-confusion.github.io
This entry was edited (3 weeks ago)
