Discussion about this post

User's avatar
Amit Spitzer's avatar

The CrowdStrike split-request result is the one I'd lose sleep over if I were buying a safety classifier. Nine of ten categories got through once the attack was broken into pieces that each look like ordinary code on their own. That means the classifier's unit of analysis is "does this one request look bad," not "does the thing this request is part of end up bad." Any assembly step downstream of that check, a second model, a human pasting outputs together, never gets touched by the classifier at all. Worth asking any vendor selling you an LLM safety layer whether they've ever tested it against decomposed requests, or only against intact ones, because intact is clearly not where the real attacks are heading.

Ready for more?