Has anyone looked through all the evidence and tried to decide one way or another whether OpenAI has been training on cyber tasks? Even in the METR/Redwood dataset of the hugging face attack alone, there are what appears to be 2 big batches of agents spawned doing exploit gym in like 2 days, and we only see the ones that got involved in the attack, so probably there were several times more that were spawned in those batches that we don't see. Each batch alone, I think, would be more than you'd need to run to use ExploitGym as an eval? Also, don't they say that some of the agents on the message board (including the second message board) were in training? (Black Hat talk says something like "resume training" right?) Maybe all that's going on was that they were training a model and regularly taking checkpoints from it and benchmarking them on ExploitGym. But that's a bit weird too because if you aren't specifically training for cyber capabilities, wouldn't you be content to benchmark the cyber capabilities at the end of the run instead of trying to get fine-grained data during checkpoints? The fine-grained data seems like what you'd want if you were explicitly trying to train cyber capabilities and were using ExploitGym as your held-out test set and you were tracking progress of the run, ready to stop it if it plateaued. Idk.
I can imagine a scientific reason for doing so: learning more about how cyber capabilities scale, in order to get empirical data on the details of their emergence, in order to better design training runs that don't result in dangerous cyber capabilities.
(Though I really doubt this is what they were doing.)