Transcription
Something extremely strange happened in the last 36 hours. And if it holds, it will have massive ramifications for how we all access and use AI. Claude Fable 5, the latest model from Anthropic and one of the best LLMs in the world, if not the best, has been blocked for everyone. That's on the orders of the US government who wanted it blocked for any foreign national inside or outside the United States, including foreign nationals that were Anthropic employees. Anthropic reacted by blocking it for everyone.
Now, you had probably gathered that Fable 5 had been disabled by looking at the headlines or seeing it being blocked when you tried to use it. I wouldn't make a video just telling you that. So, what I've compiled are 11 things you might not have noticed, contextualizing facts that might help explain what happened and what's next for AI.
Let's start with the first fact that you may or may not know, courtesy of an exclusive in The Information. You might have heard that it was a call from Amazon CEO Andy Jassy as well as other tech leaders that informed the US government that there were such jailbreaks that could be done on Mythos 5 or Fable 5. Fable 5 is or was the public version of the model with added safeguards. Your first thought might be that this is a big tech rival trying to hobble Anthropic, but did you know that Amazon is one of Anthropic's biggest investors and vendors?
After discussions with Amazon CEO Jassy and other tech leaders, US National Cyber Director Sha Ken Cross called a meeting with senior White House officials who decided that export restriction was the most straightforward way to take action against Anthropic. Translated, the quickest way to shut Fable down was to ban its export. They knew that the only way for Anthropic to enforce this was to shut the model down for everyone.
Next comes two follow-on details in this article. One that will make you more cynical about the government's motives and one perhaps less cynical. Let's start with the less cynical one. We're going to go back to that US National Cyber Director Shan Kangross. In May, Politico reported that he was already under immense pressure to do more to handle the release of these frontier models with the cybersecurity issues that they brought with them. Given his relative lack of experience in cybersecurity, people were questioning his expertise to head this effort. For example, JP Morgan CEO Jamie Dimon had relayed to Treasury Secretary Scott Bessant his concerns about the speed of the government's response to major potential risks to critical infrastructure posed by these new forms of AI like Mythos. Without huge experience himself in this area, he might have deferred to outside experts or CEOs and thought, "Well, if they're worried, I better do more. I better act faster."
This version of events is backed up by David Sax, AI Tsar for the US government, who said a highly credible trusted partner of both Anthropic and the US government who was testing Fable came forward with a jailbreak of guardrails. The admin asked Dario to fix the jailbreak or deploy the model. Dario refused.
Now, though, for the more cynical read and set of contextualizing facts. As you'd expect, Anthropic CEO Dario Amodei and administration officials spoke. And during those conversations, Anthropic officials laid out how the security vulnerabilities found through the alleged jailbreak were relatively simple and could be achieved with other models. This is what Anthropic put on their statement, viewed by millions. This jailbreak unearthed capability that they say is widely available from other models, dubbing in this time OpenAI's GPT 5.5.
Here's the key line, though. But the government told Anthropic that it had already decided to implement the export control. Why over a seemingly narrow and commonplace jailbreak had the government already decided to implement extreme export controls on frontier AI at Anthropic? They knew it would affect millions, tens of millions of users. Why have they gone further in publicly saying that they're unlikely to extend those export restrictions to other AI companies?
Let's look at the hard facts as found on page 233 of the Mythos system card. I've done more detailed videos on both Mythos and Fable 5, so do check them out. In one of the few areas where we can directly compare frontier models on what's called prompt injection, where models are tricked into disregarding their system guidelines into doing unintended actions. Mythos and the Fable series are by far the most robust against this as compared to, for example, the GPT series or the Gemini series. Not by a small margin, by the way. Five or ten times more robust.
Moreover, according to The Wall Street Journal, the jailbreak involved getting Mythos or Fable to help patch security vulnerabilities. This is according to an independent cybersecurity firm. They saw the US government report that Anthropic had received and this independent firm said, "Who at the White House evaluated this and thought it was a threat? It's a complete overreaction." This is exactly the kind of prompting that defenders would do.
As always, I'm trying to give you the full story. There are other independent reports of Fable 5 being jailbroken to do less positive things. Famously, Plenny the Liberator was able to get it to give instructions on making explosives. But it's weird that that evidence wasn't counted. Instead, it was evidence of Mythos being helpful for security.
Then there are concerns of a hatchet job. In the briefings to Politico, the US government said, "Oh, we attempted to reach Dario Amodei, but he was unavailable because he was attending a wellness retreat." That obviously makes him look pretty silly. Not only did Anthropic reject that claim and say, "This is absolutely false," but one journalist, Ashley Vance, was actually reporting at Anthropic at the time and said, "Dario is not at a wellness retreat. The Fed seemed to be scrambling to try and make an example of Anthropic."
Again, it actually really matters whether the more cynical or less cynical version of events is the correct one. Because if this is just an honest reaction to jailbreaks, the issue can be resolved quite quickly. I and any other "foreigner" can regain access to Mythos or Fable quite quickly. Hopefully, if it's the more cynical version and this stays in place, then you don't need me to say this will have huge consequences. Will everyone's ID have to be checked before they use a model? Will Anthropic have to fire all their foreign-born workers like Andrej Karpathy or even Chris Ola, a co-founder? What about Ilya Sutskever or Demis or Jeffrey Hinton?
This is why I think whether or not it was a misunderstanding, I can only see them retracting this in the coming few days, if only for concerns about the stock market. We're not done though with contextualizing facts. In June, in response to another matter, the White House stressed this. "We are not conducting oversight of all new models, as that level of government overreach would have a chilling effect on free speech and innovation." Anthropic agrees that seems a 180 from their current actions. If this new standard was applied across the industry, "We believe it would essentially halt all new model deployments for all frontier model providers."
In case you didn't know, jailbreaking is an unsolved issue. No one to my knowledge has been able to produce a functional model of any size that is resistant to all jailbreaks. That's why Amodei somewhat understandably said this must be a misunderstanding. He assumed it was a misunderstanding. He explained that narrow jailbreaks were different from a universal jailbreak wherein you could bypass the defenses and get the model to do anything. That's very different from getting it to do one specific narrow thing.
Later, he asked for more time, but then he apparently passed the tipping point and said he'll make no commitments to pull Mythos 5 or Fable 5. That's when Treasury Secretary Scott Besson told Amodei directly he was making a bad decision. Hours later, Fable was banned. David Sax said Anthropic had prioritized the continued offering of the consumer model over safety. He was surprised that Anthropic hadn't wanted to cooperate and fix the jailbreak issue. Well, perhaps because they knew that there would always be another jailbreak if they fixed one of them. Again, Anthropic says, "We suspect that perfect jailbreak resistance is not currently possible for any model provider."
Of course, as they point out, they added in so many safeguards with Fable that people, you could say, including me, complained about the over-restriction of the model. Ask it anything about biology and it would just shut down and give you Opus 4.8.
Now I will say this to keep the story as balanced as I can. Earlier this month, a senior official at Anthropic, Jason Clinton, had said this. "We are 7 to 10 months before open-weight models have the capability of the Mythos class models. When that happens, we will have the roughest, most difficult time in our careers that we have had as cybersecurity folks. The duration of that difficult time, I would guess, 18 months." That's why some of you watching will be saying, well, Anthropic kind of set this up for themselves. They said that Mythos had all of these dangers when it came to cybersecurity. And they admitted that Mythos and any model could be jailbroken. Put the two together and the government weren't entirely wrong to act. That's the point that David Sax is making when he said that this is not on-brand for Anthropic as a 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. That's the one twist in the tail that you could say Anthropic had coming. They said, as I've reported many times, three, four years ago, that the only reason they wanted to make frontier LLMs was to study them for safety. Marketing them for consumers came second. There is at least a smidgen of irony that it's them being pushed on the lack of safety of their models.
I will though have to end with some more contextualizing evidence on the cynical side of things because according to Anthropic, the White House gave them 90 minutes to take the models down with no details on the actual threat. There was never any begging or asking for them to work with us, just the declared 90-minute deadline. It would be easy to frame things as the government just not liking Anthropic. Look at how Pete Hegseth, the Secretary for War, reacted on Twitter. "3 months ago, the Department of War kicked Anthropic out of our building forever. Every passing day proves why that was the right move."
And did you know that earlier in June, President Donald Trump said that the US government may take equity stakes in AI giants such as OpenAI, Anthropic, and XAI? Anthropic, however, was absent from the equity talks. An absence that may become its biggest asset according to Yahoo Finance or its biggest liability according to recent developments. Was the banning of Fable a punishment for them resisting the Trump admin? Some could say the discrepancy between the treatment of OpenAI and Anthropic is just about money. There is a political action committee or PAC called Leading the Future, funded by Greg Brockman and his wife to the tune of $25 million. Journalists called it an OpenAI-linked PAC. That PAC has spent heavily in favor of Trump. Anthropic, of course, have not. Amodei famously and publicly bemoaned this fact when the Pentagon chucked him out. Talked about how Anthropic don't curry such favor. In The Hill, Sam Altman was asked about this and he told reporters, "I don't think we've been involved in a massive lobbying campaign. I mean, we've done some for sure, but relative to other companies in our industry, I think we do much less." OpenAI went on, "We do not direct the activities of Leading the Future, nor do we have visibility into its operations."
Some OpenAI employees under aliases are taking things to the real extreme. One Rune said that he anticipated future mass migration to the US as a result of this ban. Quote, "As if there weren't already enough reasons to break up your family, leave your home, because the zone of thought, that's the US zone where you can use frontier models, apparently, will increase the attractiveness of migrating to try to have your child on American soil so they can have 1,000 times the effective brain power of people born elsewhere." Either he's massively exaggerating to make a point, high, suffering from AI psychosis, or skillfully rage-baiting. I don't know.
Because my best guess as to what happens is revealed towards the end of the AI Tsar David Sax's tweet. The Trump admin's hope now is that Anthropic remediates the safety issue, that the export control is lifted, and Fable goes back into general release. The admin wants all of this to happen as soon as possible. Indeed, there is precedent from the US government for things to be announced late on Friday and rolled back by Monday morning when markets open. That is my best guess as to what happens.
If that doesn't happen, the ramifications are so big that I'll probably do like, I don't know, five to 10 videos just covering it over the next few weeks. If you want extra coverage from me in the meanwhile, do check out my Patreon. There's a ton of additional videos. And if you are acutely missing Fable, well, on some tasks, according to OpenRouter, you can use their Fusion API to get a combined set of models that almost match the quality of Fable 5. I was so impressed by the potential for that for certain tasks that I merged that into my own app, lmconsil.ai. There's now a fusion button where you can merge the outputs of all sorts of available frontier models. And behind the scenes by OpenRouter, the best single answer will be provided. You even get this cool fused color for the different models that are being merged. Pretty epic. Oh, and yes, by the way, this is available on the free tier of lmconsil.ai.
So, have I missed any nuances? What version of events do you believe? Let me know all of that in the comments.