TLDR version: the fine-tuned model they describe improves things, but I still don't think "improves" is good enough If you are facing an adversarial attacker then reducing the chance that they might find an exploit just means they’ll try harder until they find one that works

→ View original post on X — @simonw,