TypeSafe's Jev against four shipped prompt-injection detectors on 7,803 labeled messages and 296 conversations. It wins the curated benchmark and the newest attack set, loses two older ones, and its probabilities track…

3 points•ramoz•2 months ago•2 comments•

2 comments

spdustin2 months ago
Sorry, I closed the tab after the first two Claude-isms.

It's _so_ frustrating.

ramozabout 2 months ago
fixed it

Read the full thread on Hacker News →

Related stories