Also, the reason I ask is that we’re developing a control-style monitoring benchmark based on BashArena, and I was wondering how OpenAI is doing this. We’ll release the full new dataset of attacks within the next month, but you might be interested in seeing our early results.
Thanks for the links, this is super useful!
Also, the reason I ask is that we’re developing a control-style monitoring benchmark based on BashArena, and I was wondering how OpenAI is doing this. We’ll release the full new dataset of attacks within the next month, but you might be interested in seeing our early results.