let me get this straight.
anthropic launches fable.
it took few days for the u.s. government to force it offline.
all cuz amazon discovered a jailbreak which appears to be a standard cyber security workflow that asks the model to fix vulnerable code.
is this a giant cyber security vulnerability?
well anthropic doesn’t think so. you can do this with gpt 5.5. so why single fable out for this specifically? if you do, then should every serious model be export controlled the moment it can do useful security work?
didn’t matter cuz the admin dropped an export control order anyway.
sacks now says we told dario to fix it which is interesting because what exactly is this fix? how do you fix this?
more specifically what exactly is the regulation? are you regulating the model itself? a capability? a prompt pattern? some cybersecurity workflow?
seems to me that anthropic would have a lot to lose if this model actually caused serious widespread security vulnerabilities in the first place.
