Amazon CEO's talks with U.S. officials triggered crackdown on Anthropic models
wsj.com698 points by ls612 18 hours ago
698 points by ls612 18 hours ago
https://www.wsj.com/tech/ai/amazon-ceos-talks-with-u-s-offic...
I still am struggling to understand why they informed the government about something that is known to be an issue in every LLM. There is no LLM that cannot be jailbroken, so unless this means that we have reached the absolute maximum publicly accessible US made LLMs are allowed to operate at with GPT 5.5, this is not grounded in any sane regulation attempt. Does anyone know what limits Fable 5 has overstepped in the eyes of the government? Parameter count? Certain benchmark results? Training computer? Cause if it’s just the ability to assist with cyberattacks and being jailbreakable, there is no model previously released that isn’t equally guilty. Remember that for GPT 5.5 and 5.4, OpenAI also restricted the cybersecurity focused use under designated models, otherwise rerouting to 5.3-codex like Fable did with Opus 4.8. And both OpenAI models can also be jailbroken all the same. Basically, what was the reason to tell the government now and not with Opus 4.5 or GPT 5.4? sama has been doing the rounds with apocalyptic predictions… I submitted separately, but this Axios report has some details that call a lot of the speculation in this thread into question, i.e. that this wasn't much of a "jailbreak" at all and that it's not Anthropic-specific - the White House intends to generally regulate Mythos-class models (whatever exactly that means): Between the lines: The government's response "seems way out of line with what's actually in the research report," Luta Security CEO Katie Moussouris, who Anthropic shared the Amazon report with, told Axios. Moussouris said the researchers were able to find security vulnerabilities by asking questions normal defenders would ask AI, which is exactly what the model was intended to do. An administration official told Axios they do not view other models as national security threats because they do not surpass the bar that Mythos set. Anything at Mythos level or above would need to go through the administration to ensure the government's national security apparatus is hardened enough, the official added. https://www.axios.com/2026/06/13/anthropic-amazon-white-hous... > the White House intends to generally regulate Mythos-class models (whatever exactly that means) This is not at all surprising. And I hope people don't make the mistake that it's a "this administration" problem. It was obviously from the early days of these LLMs that the shoe was going to drop and we (as Joe public) would not retain access. I mean that once ChatGPT3 dropped it was clear there was some level of functionality at which we would be denied further access. The only carve out will be as per older technical innovations the US is more concerned with foreign national access than US citizen access at home. I don't remember the details with encryption but it was basically you have to ship a breakable version for the rest of the world, and you generally sometimes ship a backdoored version. And Anthropic is more concerned by what they are asked to do to US citizens than the broader group. Same story with encryption, CPUs, GPUs, blah blah blah. Yet unlike CPUs/GPUs, there's currently zero way to lock down who has access. Giving access to 'citizens', with the current way the Internet operates, is absurd. One back door into a desktop, workstation, and 'validated citizens' are now 'hackers from where-ever'. >and 'validated citizens' are now 'hackers from where-ever'. Yes, because knowledge is power, and information is meant to be free. That’s a terrible way to create AI regulations If they actually cared about this issue we’d have predictable laws and regulatory bodies that let companies actually plan There’s a reason royal fiat doesn’t lead to healthy economies. It’s just confusing and chaotic. It’s not clear why anyone would invest in a new model now. Then the next administration comes in and instantly, by fiat, they decide to lift the ban. The market just gets jerked around with no ability to plan long term investments. It’s a great way to regulate if you’re corrupt. When the rules are opaque and arbitrary, there’s a lot more room for corruption. [flagged] Whether or not you agree with how US laws are drafted, this administration has no logical foundation for anything it does which is a massively different and worse problem by orders of magnitude. This administration runs on whims. This is horrifying and there is real harm in this we have yet to see the full repercussions of. You are biased, previous administration war on crypto was worse IMO. The attacks on private banking for companies dealing with crypto and 0 laws by the SEC. This is a fact regardless if you like/dislike crypto. The lack of a logical foundation isn't the novelty. The whole system has run on whims and backfilled reasoning for a long time. That's the problem. If it had always been the rule of law until now then we would have an apparatus set up to impose checks and balances and accountability on government officials, but because those things have so atrophied from continuous contempt and neglect, no one knows how to demonstrate that what Trump is doing is wrong without also conceding that half of what the government has been doing for decades is wrong. But they also don't want to stop doing those things and therefore have rather a dilemma. Of course, that's assuming you actually demand logical consistency. If you don't care about that you can do whatever you want -- which is kind of the trouble. In countries other than the US, most regulatory bodies are outside the government for exactly that reason - to take the power away from the political elite, whilst continuing to ensure safety and reason come first. The new law the US is proposing here, is the exact opposite. A kingly appointed adjudicator to decide things. Not that I'm ever one to support anything this regime does but I'm kind of okay with them pumping the brakes on this until we really get a handle on what the The USG has limited capabilities on technologies from GPS chips to thermal imaging with "national security" implications for a while and now they're doing it but it seems people don't like how ill defined "Mythos-class" means. Would it be better if it was some %X on some benchmark that the frontier model peddlers could just limbo under to make it "acceptable" for release? Do we just accept that jailbreaking will never be prevented? The part of all this I do have a problem with is the national state cybersecurity cat-and-mouse this kicks off. Will the US tech landscape have enough time to safely get a "Mythos-class" model to harden itself before China releases or leverages a "Mythos-class" cyber munition? "pumping the brakes" would be fine. This is slamming to a full stop on a crowded freeway and causing a three car pile-up. Warning and advanced notice are the difference between regulation and tyranny, and in this case we're just getting tyranny Same problem as always. This administration never figured out that how you do things matters. They love the drama of the crash more than actually implementing functional policies. Given the current climate I'd be inclined to declare "tyranny" also but in this case I think given the degree of potential damage the slamming on of brakes is warranted when the alternative is, to strain a metaphor, going full speed off a cliff at relativistic speeds. Yeah, we have a lot of critical infrastructure connected to the internet. Based on the trend the last few weeks, I expect major cyber attacks this year. I expect that to happen no matter what we do (since the open source models are rapidly catching up), but gating access to the frontier models for a while sounds like a reasonable precaution — as annoying as it is to me personally, to be deprived of such shiny toys! Fable is a massive step up and I didn't expect it public for another month or two. Something tells me we'll get it back in a few weeks though. > and in this case we're just getting tyranny You expected different with this administration? Of course I expect the government to act better than this! But I am not so naive as to assume my expectations will be met. I have no insider information so this is all appreciation, but: When it comes to legislative things, there is pretty much always a timeline in which to become compliant. I do wonder if there was opportunity to give warning etc. but Anthropic decided to perform an immediate full stop deliberately causing the metaphorical three-car pileup, because the more painful for the users, the more pressure from the people there will be on the government to undo this. See also: those painfully annoying cookie banners that are malicious compliance in the most irritating way possible, which GDPR does not require, in order to make people think GDPR is dumb. > The USG has limited capabilities on technologies from GPS chips Are you referring to Selective Availability? That ended decades ago. Selective Availability accuracy restrictions ended decades ago, but GPS technology is still subject to various military and export-control restrictions. Not selective availability. COCOM Limits that prevent a GPS chip from operating above a certain speed and altitude. It’s funny because it’s just (relativistic) math. It would cost a couple hundred bucks to roll your own with no restrictions. > That’s a terrible way to create AI regulations This administration doesn't do regulations, its extortion. Same as the tariffs. Just grease someone's palm and then the vague restriction is lifted. I still can get de minimus from China no problem, as long as it’s Ali express. I wonder why? When anthropic answers that question, we will have access to fable again. In a parallel universe where we have Biden (or Democratic Party) administration, how different do you think the regulations / approach would be for this fast moving and unpredictable technology? It’s hard not to see this ban as being motivated by retribution for refusing to use the models for spying and autonomous warfare. Probably using the rule of law in some way? Talking about it in public? Legislating? You know... government type stuff? Like Biden did for crypto? Oh wait, no he had a backdoor war using the banking system and refused to enact new laws. Refusing to enact new laws around a thing most people don’t like, don’t want, and don’t care about (oh and is used for scams often) is quite different than a secret back door war. They at least wouldn't depend on how extensively you publicly glaze the President. There is not a single chance this would have happened under that admin. Not one single chance. It doesn’t really matter what party does it The ideal case is a statutory agency with regulatory authority that sets very clear standards for what model capabilities can and cannot release. Those are set ahead of time and well known by frontier model providers. Most normal regulations are managed through the administrative procedures act process. That’s a legal requirement that involves deliberation and public comment. I’d argue you could pretty easily enumerate most capabilities that have been obvious concerns for a while. For example, cyber security. This structure can last decades and reassure players they can operate in the market without rules changing suddenly without warning. Some kind of sudden, temporary action like this export control tool is legally fragile. Even if sometimes necessary in exceptional cases. But if the administration sees this as a permanent way of working, they won’t be helping anyone (but maybe themselves through grift). If the administration truly cares about functional regulation (which maybe they don’t) they need a sturdier legal structure that lasts past Trump. Not flimsy edicts that change with the wind I wholeheartedly agree with what you’re saying in general. I do wonder though, given how rapid advancements in AI are occurring, if even an agency with statutory authority would be able to establish a predictable regulatory environment, let alone do so while maintaining a lengthy public comment period and a whole of government approach. There are obvious flaws with the current administration’s approach to, well, almost everything. But I’m not sure if this is even a tractable problem with the governance structures we have been employing over the last 50 years. Nothing being talked about with Mythos wasn’t a known AI risk 12 months ago. Those rules could have been established to guide frontier labs. But yes crazy things happen. Maybe it won’t catch everything. The right answer are giving the govt / this agency explicit legal, short term model pause capabilities to let the rule making process happen if something completely out of band happens. Or let the agency study/approve models prior to release. Not sudden, unexpected application of export laws. Yet in this case, for Fable, cybersecurity risks have been well know for some time. A rule created years ago when we knew this would happen could have given frontier labs and the market predictability. They probably would have been in line with Executive Order 14110, the Biden administration's detailed description of a principled approach to regulation of the AI industry. It would have been aligned with the Trump administration's stated goals as well, but a coalition of rich VCs successfully bribed him to rescind it as one of his first acts in office, because the primary principle of Trumpist government is that people who pay Donald Trump a lot of money get what they want. Why amazon? I bet the three letters had a hissy fit field day worrying that their expensive hancrafted zero days would evaporate and software would get more secure. So, the government is throwing a wrench for the NSA Interesting. Hope there is any clarification on what "Mythos level" is and why 5.5-cyber doesn't arise to it. Any metric I could come up with (parameters, pre-train compute, benchmark scores, etc.) seems somewhere between imperfect and utterly nonsensical. Pure speculation, but GPT-5 series models including the new 5.5 pre-train appear far closer to Sonnet than Opus or Fable in pure parameter count, so maybe that's it, but the "they do not surpass the bar that Mythos set" line sounds more like there is a believe that Mythos/Fable are more capable in cybersecurity tasks, whereas the data [0] doesn't seem to bare this out. I did not do any cybersecurity assessment of Fable 5 myself, partly due to personal reasons that make that something I'm abstaining from, but my coding evals showed that while task adherence and assessment wise it was neck and neck with 5.5, the task inference was a major jump again (something prior Anthropic models tended to already do incredibly well on) and while that makes it a far better model to work with for UX experiments, I don't see how that translates to cybersecurity, along with the aforementioned publicly available evals by AISI. Seeing as neither Mythos nor GPT-5.5 had been pre-trained with a particular focus on cybersecurity, this would have to mean any model that benchmarks better than GPT-5.4 or Opus 4.6 on these tasks cannot be used by None-US-Citizens. If such guidance isn't enforced for all US labs, I think that's irrefutable evidence that this isn't about cybersecurity or "the bar that Mythos set"... [0] https://xcancel.com/AISecurityInst/status/205458976317312633... Firefox bugs found per month, actively advertised as a sign of how powerful Mythos is: https://external-content.duckduckgo.com/iu/?u=https%3A%2F%2F... I am, thus far, not aware of 5.5-Cyber managing anything similar to "Project Glasswing" That said, the government also knew about Mythos since Project Glasswing was announced... April 7th, two months ago, so if they wanted to block a public release, they had more than enough time to do it in an orderly way. And basically every sign that Mythos is well above the previous baseline was pretty publicly known by early May, when we started getting stuff like the Firefox bug reports. I can see an argument that Mythos is just barely a "cut above" enough to regulate, but I cannot see any argument for doing this by a fiat order three days after the release. They literally asked for it. Two days ago Amodei wrote an essay urging the government to regulate them. He explicitly cited Mythos, as proof that frontier AI has acquired autonomous hacking capabilities that threaten critical infrastructure and national security. A third-party demonstrated that it was possible to jailbreak the safety measures of Fable to access the raw Mythos abilities. Abilities which Anthropic say are too dangerous for the public. Edit. From David Sacks: Cynically: this is an attempt to quash open source or discount model competition through regulatory capture. I'm sure it's also a step towards requiring id and limiting access for us plebians to real power and keeping it for maintaining or growing power of those in charge. It's all an excuse to give us a Westworld season 3. Probably a better example out there.. > A third-party demonstrated that it was possible to jailbreak the safety measures of Fable to access the raw Mythos abilities. Abilities which Anthropic say are too dangerous for the public. Pressure test this assumption before getting behind this position. I will certainly revisit it as more information comes out, but is it your contention that Anthropic solved jailbreaking with Mythos? What you claim contradicts Anthropic’s statements. I assume that is the contention. That is a strawman. My contention is what you just implicitly acknowledged - there is not information put out yet to validate the quoted claim. There are claims to the contrary, as well, from Anthropic themselves. In the absence of information, maybe it’s better to ask which claim is more extraordinary. That, A. Anthropic solved the llm jailbreak problem with mythos (despite no claim to have done so on their part) B. That a full jailbreak of mythos is possible. That’s not what the claim is though. Anthropic’s claims are as follows if you read their post: * this is not a universal jailbreak method * the jailbreak affords you the same capabilities you get already with other models, not Mythos. In this situation it’s which party do you trust more and history would suggest this administration is very playful with the truth, especially when it comes to economically damaging the company that’s become their political enemy There is not an absence of information. There is information, from Anthropic, concerning the jailbreaks that motivated this action, that directly contradicts the statement. There is just an absence of information backing the statement I responded to. I find it so odd this is apparently so contentious a take. The existence of a jailbreak free llm in 2026 is extremely contentious to me. You can argue about the specifics of this exact jailbreak, but generally pliny and amazon both reported mythos jailbreaks in <7 days. It seems very reasonable to expect that a well funded state actor could achieve better results given significantly more funding, determination and most importantly unfettered access. Nobody here is claiming fable is jailbreak free. Not anthropic and not in this thread. This was known before launch. The question remains one of degree and capabilities.
Topfi - 17 hours ago
themgt - 14 hours ago
rustyhancock - 2 hours ago
b112 - an hour ago
x______________ - 15 minutes ago
softwaredoug - 13 hours ago
aqme28 - 12 hours ago
_heimdall - 10 hours ago
Kina - 4 hours ago
mlrtime - 2 hours ago
AnthonyMouse - 3 hours ago
shakna - 2 hours ago
VectorLock - 11 hours ago
handoflixue - 8 hours ago
AnneTrotter - 7 minutes ago
VectorLock - 6 hours ago
andai - 2 hours ago
voidfunc - 7 hours ago
handoflixue - 6 hours ago
chias - 3 hours ago
thfuran - 10 hours ago
rcruzeiro - 8 hours ago
VectorLock - 6 hours ago
withinboredom - 4 hours ago
tapoxi - 12 hours ago
K0balt - 6 hours ago
andsoitis - 11 hours ago
digitaltrees - 9 hours ago
hilariously - 9 hours ago
mlrtime - 2 hours ago
vineyardmike - an hour ago
ceejayoz - 11 hours ago
b--l - 5 hours ago
softwaredoug - 8 hours ago
derektank - 8 hours ago
softwaredoug - 26 minutes ago
SpicyLemonZest - 11 hours ago
warumdarum - 3 hours ago
Topfi - 13 hours ago
handoflixue - 8 hours ago
irthomasthomas - 13 hours ago
https://darioamodei.com/post/policy-on-the-ai-exponential "Mythos Preview scrambled the global cybersecurity landscape. But its broader significance is that it proves beyond doubt that AI models are now tools of global and national strategic consequence."
"The government should have the power to block or deter deployment of the model if it is determined, in light of third-party assessment, to present unacceptable risks. This power must be scoped to the above four specific risks and there must be protective measures against political favoritism or arbitrary decisions"
— A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused.
— In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the USG believe; nor is that kind of minimizing language consistent with Anthropic’s brand as the AI safety company. It’s difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not “serious".
crowbahr - 11 hours ago
sanex - 11 hours ago
sigmarule - 12 hours ago
irthomasthomas - 12 hours ago
apstls - 10 hours ago
sigmarule - 12 hours ago
drawnwren - 11 hours ago
vlovich123 - 10 hours ago
sigmarule - 10 hours ago
drawnwren - 6 hours ago
s1artibartfast - 5 hours ago
