The Commerce Department directive hit Anthropic just days after Fable 5 and Mythos 5 went live, forcing the company to cut off every user worldwide rather than attempt to segregate foreign access on the fly.
The admin is going to keep leaking all sorts of details that make it sound, to a casual reader, like this was a reasonable decision.
But there has been zero indicating that domain experts at CAISI or NSA were involved - all reporting points to "senior White House officials."
Senior White House staff drove the move with selective leaks afterward, leaving CAISI and NSA experts on the sidelines and creating the impression of a rushed national-security call.
No public evidence has surfaced on the reported jailbreak method or any licensing path forward, so affected developers and researchers have no timeline for restored access.
Many users criticized the Trump administration's AI licensing regime via export controls as arbitrary, inconsistent, and shaped by favoritism plus inexpert decision-making without expert input.
To be clear, I think it's quite plausible to imagine a reasonable case that "the Fable launch was rushed."
Threat modeling, cost-benefit analysis, etc. are hard.
Less likely is a good case that "this was clearly bad *and* also 5.5 was (and soon 5.6 will be) clearly good."
@Miles_Brundage Perhaps that’s true in the future, it was the Amazon CEO who raised the alarm. You kind of have to respond if you’re the government, if one of the largest investors in Anthropic that also is tightly coupled with their compute gets spooked
The closest that there has been to such an indication is the claim that the jailbreak research was shared with "security researchers."
But there is no indication that this had influence on decision-making + the 1 independent researcher who has been quoted concurs with Anthropic
@Miles_Brundage Caincross was involved, he's the National Cyber DIrector
@Miles_Brundage +1, vibes
The admin is going to keep leaking all sorts of details that make it sound, to a casual reader, like this was a reasonable decision.
But there has been zero indicating that domain experts at CAISI or NSA were involved - all reporting points to "senior White House officials."
(and there's no evidence that even that better version of a case was made, and every reason to suspect favoritism + inexpert decision-making given *gestures broadly*)
@Miles_Brundage Agree on all of this and the hypocrisy...but from a safety standpoint, the govt pushing back on model releases with safety worries is not per se crazy, yes? I agree "lack of expert knowledge", but "Amazon comes to USG and claims the mythos v fable distinction isn't ironclad"...
@Miles_Brundage For sure. I think problem is anthropic (correctly!) claimed they had a step jump in cyber capabilities, did a month of work on this, released a model with limits to make it safe on these grounds, and their own partner (unlike in 5.5) told govt high-ups the limits are imperfect...
@Miles_Brundage This seems like a soft touch in response to Anthropic saying "Yes, Fable is an incredibly dangerous national security risk, and no we can't stop full jailbreaks."
@Afinetheorem Agree it's not per se crazy - I just have low confidence that the subtleties of different types of jailbreaks, threat models, etc. were considered + that it's justifiable to do this and not (e.g.) pull GPT-5.5 too
@Afinetheorem Not sure if we disagree here - like I don't think it's impossible some people earnestly thought this was the right call. I just think they are likely missing key context + mistaken, + someone bears some responsibility for not looping in those with more context before deciding.
@Miles_Brundage That distinction matters a lot
Right, under risk frameworks of the DoW, less intrusive methods and assessments could be made to carve out authorizations for critical staff at Anthropic who are foreign nationals, to effect the stated goal of addressing an addressable risk to nat security. Highly contradictory action vs intent.
@Miles_Brundage so, podcasters
@Miles_Brundage @davidmanheim admin talks while the pros shut up. classic shit.
@Miles_Brundage It did sound like Dario was told something, waved it off, and then was confused when USG reacted, in the Anthropic announcement “we believe this is a misunderstanding.” Seems more Dario’s misunderstanding…
@Miles_Brundage Yes, indeed. Or put otherwise, "done like everything in the Trump admin", hah! But I wish folks both in US and outside would understand "*no* govt will let you release whatever frontier AI you want to whatever customer you want wherever they are" will be true very very soon.
@natolambert @Miles_Brundage National Cyber Director was involved
@Miles_Brundage Yep. The ugly cases are when 5.5 ships under one eval gate, 5.6 under another, and nobody updates the rollback owner or incident threshold.
@Miles_Brundage If a Fortune 500 pushed a high-risk model update without signoff from safety, red team, and the release owner, audit would have a field day.