@code_star There's also a lot of mapping you can do with clever thinking around negative spaces and refusals. I used to do this a lot with hitting ai research guardrails in o3, when I could identify ideas where I could consistently tap the electric fence.

It’s almost certainly because they did some kind of A/B testing.

My guess is they defined frontier models development so broadly to transparently refuse it would essentially stop working.

I could also see them thinking that a transparent refusal could be worked around with volume of requests.

@code_star I think it's more likely that ANT has assessed that most x-risk categories are not "core" use cases of Claude. I don't think anyone is really worried about needing to compete for the "LLM-using biologist" market, and so you can just take an axe to that side of the model.

@code_star Ab testing is the go to excuse for every weird rollout decision now

lowkey does make sense as a loophole though