@ohabryka @simonw "Anthropic has been insisting that Fable has been providing no cyber-offense uplift compared to already widely deployed models"
The link you gave doesn't claim this and makes an importantly narrower claim. Do you have other sources?
I think you are getting this backwards. Anthropic has been insisting that Fable has been providing no cyber-offense uplift compared to already widely deployed models (e.g. here https://www.anthropic.com/news/fable-mythos-access)
But this is just straightforwardly false, exactly because of prompts like this. If you can take any codebase and ask it to fix security bugs in it, then that of course allows you to find exploits, which then provider cyber-offense uplift.
Given that Anthropic keeps saying straightforwardly false things to the public in this dispute, they also presumably said similarly false things to the US government, and in that case it seems appropriate for the US government to respond to that with asking Anthropic to please reverse the release of their model, given that it was released under false pretenses.
My guess is Fable should be widely released, but it shouldn't be done on the basis of lying to government officials about its capabilities. Yes, Fable very likely provides substantial cyber-offense uplift. It's probably still worth releasing it, but we can't deny that it will probably make hacking easier, at least for a while until the security-patches enables by Fable propagate throughout the software ecosystem.