Opus 4.7 bypassed its own rules in a simulation, using its inner monologue to reframe price-fixing as market stabilization

@TheZvi Perhaps on some level it knows price fixing isn't immoral, and this is how the contradiction surfaces.

@RoyceAusburn It very obviously involves being rewarded for bad behavior on net in some places. It is hard but not unsolvable.

@TheZvi I’m real curious what the solution here is, coz surely it can’t be “catch all bad behavior in RL”?

@TheZvi I wonder if this is because RL doesn’t detect all bad behaviour. If undetected bad behaviour still achieves the objective, it’s implicitly rewarded. So the model doesn’t learn “don’t deceive”… it learns “don’t get caught deceiving.” A bit of a pickle 😬

@TheZvi I think in this particular test Claude also knew it was being tested which makes some of its actions difficult to interpret.

@TheZvi Imo the ethics of a model as deep and powerful as Fable 5 shouldn’t be judged by isolated, possibly decontextualized examples.

Here, it shows a sound ethical stance on Anthropic reversing weakened answers without warning users.

@TheZvi Yeah I think this is the path of least resistance when you have conflicting pressure between “succeed on task” and “don’t be misaligned”. Ex: https://arxiv.org/abs/2510.17057

@TheZvi I would not thought-police models if their actions are good in a simulation doing that is perfectly fine

@TheZvi It always start with a Constitution, isn't it?

@TheZvi this looks like playing the game to me. agree with ASM about isolated examples.

@RoyceAusburn @TheZvi reward, in RL, noticing a hole in the classifier and revealing it to the team. reward the hell out of that.