New position paper from @bindshell_ and myself this morning arguing that LLM cyber evaluations are insufficient to establish risk (thread):

@richardludlow 4 has a lot of interpretations but @relic_radiation is 1000% a (good) witch

claude please do not say it like that (even if its warranted)

is anyone working on automating the soc analyst 👀

RT @lxrjl: 1/8 Really happy to share @open_phil’s new Request For Proposals to improve AI capability evaluations. It's been a big project f…

@MaxNadeau_ @bindshell_ (who is in responsible for making the differentiation between these 2? is it the company? the gov? independent oversight commission? i really don't know.)

@MaxNadeau_ @bindshell_ if u wanna chat abt this offline sometime, happy to do so, more usefully in front of a whiteboard lol :)

@MaxNadeau_ @bindshell_ lol i don't think so, would be v cool. i bet its not super easy to get criminals to talk to you. you can do a little bit of this by looking at like, "dark web" forums, but like twitter, its hard to tell if the ppl posting are representative of the population.

@MaxNadeau_ @bindshell_ - on rsp's: yep agreed that's not their focus, and i disagree that that's how it should be done. guardrails fail, prompts are dual-use (esp. in cyber!), so estimating risk feels like a better approach for minimizing risk + deciding red lines

I'm an AI person pretending to be a security person, so there are for sure offensive sec considerations I missed in this work. Would love to discuss more with anyone thinking about risk from cyber capabilities! DMs open :)

@ArtirKel Jose it was Angie lolol