This seems important.
I wonder if there is overlap with cyber security refusals… Could cyber security and model safety be conceptually related and that cause sandbagging on safety tasks? Model learns to avoid cyber security tasks (because they could be used for harm/exploits) and maps that to avoid working on model safety.
🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision






