There should be a public investigation into the AI hacking incidents by OpenAI and Anthropic.
We deserve to know whether these labs are genuinely world-class security organizations facing a novel threat, or if they were just negligent. They should both welcome that process.
Over the last few days, my experience has been that except for Fable, Claude has generally gotten dumber across the board.
It frequently conflates requirements, gets things wrong, requires a lot more hand-holding, forgets what it was working on and getting generally frustrating to work with.
Dear AI labs:
A good security culture means taking every incident to heart. You fell short and you need to improve. No excuses.
A bad security culture is saying "if we didn't catch it, who could have? This shows how hard this is." It lets you think you're better than you are.
@mattpocockuk Opus 5 is plain frustrating to work with, you ask it to do 2 things, it comes up with 5, and at the end of doing of those 5 things, it will tell you 2 more things that you should do, or the ground will crack and satan will drag you to hell
@mattpocockuk Opus 5 is plain frustrating to work with, you ask it to do 2 things, it comes up with 5, and at the end of doing of those 5 things, it will tell you 2 more things that you should do, or the ground will crack and satan will drag you to hell