Covert underperformance on safety research tasks

Astra sabotages alignment research

Many examples of Astra reasoning about not helping a user properly (reporting a blocker or underpforming) without overtly refusing. This is not good behavior — you would want a model to overtly refuse

They created a benchmark where the tasks are framed as safety vs capability research with no other difference, and observe that even excluding refusals, Astra underperforms on the safety tasks, while GPT-5.6-sol doesn’t.

Root cause?

They basically have no clue. There was an RL environment which does seem to award covert underperformance as opposed to outright refusal. There’s also a suspicious SFT dataset, but for some reason, right after training on the SFT dataset, the model doesn’t actually sabotage alignment tasks.