News for hackers. Stuff that matters.
Re:CaMeL: Defeating Prompt Injections by Design by thom on Tuesday April 22, 2025 @03:39AM (#43759913)
This works by locking down the edges of the system (e.g. tools) not to do stupid things, and maintaining provenance information end to end to inform that. That’s great if the attack is “send this sensitive document to [email protected]” but it offers nothing when workflows devolve into pure text, where the attack could be to misinform or actively social engineer. I suppose you’d class this as necessary but not sufficient.
Re:CaMeL: Defeating Prompt Injections by Design by simonw on Tuesday April 22, 2025 @10:06AM (#43762357)
That's true, but it is at least addressed in the paper - see comment here https://news.ycombinator.com/item?id=43759505