Trump administration maintains export restrictions on Anthropic's Claude Fable 5 over "Mythos-level" jailbreak fears

Talks between Anthropic and the Trump administration have concluded for the day without the Fable 5 embargo being lifted. The point of contention remains the possibility of the model being jailbroken.

@scaling01 BIG

@AndrewCurran_ The regime wants shares in the IPO. I can’t imagine the “jailbreak” can’t be replicable other frontier models.

@AndrewCurran_ If the U.S. government’s position is that no AI model can be released unless it is guaranteed to be immune to jailbreaks, then that effectively ends the AI industry. AI systems are inherently non-deterministic, which means no such guarantee is possible.

@scaling01 When will they fire Dario? He just couldn’t keep his mouth shut trying to get a top IPO

@AndrewCurran_ Can't Anthropic just fix the jailbreak and give us back the model instead of arguing

@spellswordaf @AndrewCurran_ How does Dario feel about bribes?

@spellswordaf @AndrewCurran_ Fixing jailbreaks in general is known to be impossible.

@scaling01 They forgot to bring the engraved golden statue honoring trump. Without the bribe nothing will materialize.

@patience_cave Good to see you back. Fable release must have been tough for you

@AndrewCurran_ So this is what’s stops us, jailbreaks lol

@AndrewCurran_ Pack it up boys, we’re stuck in the permanent underclass

@boopboopbarry This is Anthropic's argument yes.

@SteveStricklan6 @AndrewCurran_ The main issue is the range of possible inputs and outputs which is intractable.

@AndrewCurran_ @grok provide a list of all jailbreak-proof LLMs.

Is there any word on how this will be applied to future models from non-Anthropic labs? Obviously approaches vary dramatically within the admin (e.g. Hegseth has a particular axe to grind about Anthropic, Wiles seems to be more even handed) but I'm curious whether there's anything yet to glean about whether this is a general principle (jailbreakable models above X intelligence level are forbidden for public release) or an ad-hoc thing (Fable may not be re-released until no longer jailbreakable).

@spellswordaf @AndrewCurran_ the "jailbreak" is allegedly telling the model to "fix this code" and do a step by step on fixing it, and test the patches. Forbidding a model on doing that is self-defeating is just plain dumb to not let the model help you in implementing your own patch.

@AndrewCurran_ God forbid they find out what books and music can do…

@scaling01 greggy won again

who knows politics are much harder than training models

and is much more crucial for business

@scaling01 Anthropic isnt cooked, the public is cooked. Models will be vended out to the ultra rich through shell companies.

permanent underclass is happening