Post #4238213

@jerry@infosec.exchange I'm not surprised, machine learning is really good when the search space is large (all the possible interactions with the Mastodon software), but the expected outcome is small and easy to measure (unexpected privilege escalation). This doesn't tell anything about Claude or any other LLM, this fuzzing is really expensive and currently only affordable because it is subsidized by investor's money (including a portion of everyone's government-backed retirement fund in the US).