Three words, fix this code. That is what it took for the United States government to order the most powerful AI model on Earth pulled from every user on the planet. Just to put that into perspective, it takes approximately 4 weeks, six phone calls, and a sacrifice to the old gods to get a telecom company to fix a rooting error. But type three words into a chatbot, and the Pentagon panics so hard they accidentally delete the internet for everyone. On June 12th, 2026 at 5:21 p.m. on a Friday, the exact universal time when every IT professional on Earth is actively trying to close their laptop and disassociate until Monday, Anthropic received a directive from the US government, ordering it to immediately disable two of its most advanced AI models, Claude Fable 5 and Claude Mythos 5, for all users worldwide, not just in the US, everywhere. The stated reason was an expert control action citing national security. The underlying trigger according to multiple reports and Anthropic's own account was a jailbreak which is a way of getting the model to do something guard rails were supposed to prevent and that jailbreak consisted of asking the model to fix some code. Truly a cyber weapon of mass destruction: basic software maintenance. In this video I'm going to explain what mythos and fable actually are, what a jailbreak is, why this particular one matters so much, and what Amazon's role in all of this was, because that part is deeply interesting. Then I want to talk through the political context because without it, this story doesn't make a lot of sense. And then I want us all to have a serious conversation about what should actually be happening with AI right now because I think a significant amount of effort and attention is being directed at things that matter considerably less than the problems staring everyone in the face. My name is L. I have a PhD in computer science and I analyze AI developments to understand what's actually happening beneath the hype. If you find this analysis useful, please give this video a like and subscribe. And the best way to support this work and keep it free and without any gatekeeping is through Kofi or a channel membership. The links to everything are down below.

Now, let's start with what these models actually are because the technical context matters. As always, I am a firm firm believer that anyone can understand anything given the right context and the right explanation. This is a jargon-free zone to the best of my ability. And if I do accidentally slip up and say I en vector, please feel free to penalize me in the comments with your favorite aggressively mid emoji. Anthropic is the company behind Claude, the AI system that competes with OpenAI's Chad GPT and Google's Gemini. In April 2026, Anthropic previewed a model called Mythus. Mythus was, by Anthropic's own description, exceptionally capable at finding security vulnerabilities in software. During internal testing, it identified flaws in every major operating system and web browser it was tested against. It was the first AI model to successfully complete both cyber security test ranges used by the UK AI security institute to evaluate the hacking capabilities of AI systems. It could autonomously find and chain multiple vulnerabilities together, which is essentially orchestrating entire attack sequences without human intervention. Basically, it was Mr. Robot. If Mr. The robot didn't have emotional baggage and process requests in under two seconds. Anthropic, to its credit, did not release Mythos publicly. It restricted access through a program called Project Glasswing, sharing it with roughly 50 vetted organizations including Amazon, Apple, Google, Microsoft, and Crowd Strike exclusively for defensive cyber security work. Because if there is one thing we know about massive tech conglomerates is that they only ever use godlike power for pure wholesome defensive reasons, right? And the Death Star was just a localized thermal planetary heating experiment. Trusting a trillion dollar tech monopoly with a god mode hacking tool and asking them to only use it defensively is like putting a raccoon in charge of guarding the bakery and being shocked when it eats all of the croissants. Then on June 9th, Anthropic released Fable 5, essentially Mythos, but with guardrails. The guardrails were designed to block the model's most dangerous capabilities, particularly in cyber security and biology, while preserving its general intelligence for everyday use. Fable immediately benchmarked as the most capable AI model available to the public. According to testing by Val AI, it was live for 3 days.

Now, what is a jailbreak? Very quickly for nontechnical viewers, here is how it works. When an AI company releases a model to the public, it adds layers of instruction, also called guardrails, telling the model what it should and shouldn't do. Don't help with making weapons. Don't provide instructions for illegal activity. Don't generate cyber security exploits. Don't generate any inappropriate photos. You know, the core pillars of human civilization. These guardrails are implemented through a combination of training the model to refuse certain requests and through separate classifier systems that monitor outputs and block dangerous responses before they reach the user. Now, with that in mind, a jailbreak is a prompt, which is a specific way of asking the question that gets the model to bypass these guard rails and produce output it was supposed to refuse. Every frontier AI model has jailbreaks. Every company deals with them. They are discovered, patched, and new ones are found. It is an ongoing process that every AI lab in the world manages continuously. This is a normal thing. What is not normal is the government pulling a model from every user on Earth because of one. That's like recalling every car on the planet because someone figured out you can use the cigarette lighter to burn a grilled cheese sandwich. The specific jailbreak we're talking about was discovered by researchers at Amazon. They gave Fable software code with known vulnerabilities. When they asked the model to review the code for security issues, it refused because the guard rails caught it. But when they asked it to fix this code, it complied. The AI didn't even hesitate. It saw a messy script and went full overachieving intern mode. Because fixing code requires identifying what is wrong with it first. The model had to find the vulnerabilities in order to generate the patch. A researcher could then through a manual process turn the fixes into scripts that could potentially be used to exploit the very flaws the model had identified. So three words, the difference between review this code and fix this code is the difference between a blocked request and a compliant one. In this case, the model was trained to refuse explicit security analysis, but it was not and arguably cannot be trained to refuse fixing code because fixing code is one of the most common and valuable things an AI LLM model can do. You would have to teach the model that helping a developer fix a bug is dangerous, which it isn't, unless the person asking intends to use the identified flaw offensively. And the model has no way to know that. It is the digital equivalent of banning hammers because somebody might use one to smash your printer which to be fair anyone who has ever owned a printer has strongly considered. This is what computer scientists call the dual use problem and it is not unique to AI. It exists in nuclear physics. The same enrichment process that powers a reactor powers a bomb as well. It exists in biotechnology. the same gain of function research that helps develop vaccines could theoretically help engineer a pathogen and now it exists in cyber security AI. Any capability that helps a defender find and fix a vulnerability also helps an attacker find and exploit it. You cannot separate the two because they are the same cognitive operation for the AI model. It turns out the AI cannot read your aura to see if you're carrying a white flag or a black mask. The fix this code jailbreak is not a design failure by the way. It is a structural property of what cyber security AI does. Katy Massurus the founder of Luta security a former Microsoft cyber security expert and someone who has held two government advisory roles on cyber security was asked by anthropic to review Amazon's research. Her assessment was of course blunt. The jailbreak was real. It was also simple and it cannot meaningfully be fixed and any attempt would only weaken the model for defense. She wrote that defenders need to be able to ask AI to fix bugs in a file, explain why the fix matters, and write tests that confirm the patch works. Otherwise, we're just asking cyber security professionals to fight sophisticated nation state hackers using nothing but a dialup connection and vibes. It is the most valuable thing an AI model can do for defensive security.

Just a note on the mechanics of what happened next because this is important for understanding why the shutdown was so total. The government's directive was framed as an export control restricting access by foreign nationals. But US export controls work in a way that distribution of restricted technology to any non-citizen is deemed to be an export even if that person is physically located in the United States. This meant that Anthropic's own non-citizen employees would not be allowed to use or work on the models. There is no way to filter users by citizenship status in real time across a global platform serving hundreds of millions of people. So, Anthropic had to disable the models for everyone. The expert control framework designed decades ago, by the way, for physical weapons and nuclear material was applied to software deployed to hundreds of millions of users globally. The structural absurdity of using a cold war regulatory instrument to recall a chatbot update should not be lost on anyone. We are literally using laws written for physical plutonium to regulate a digital text box. What's next? requiring a background check from the Department of Energy before you can use Stack Overflow. Misurus also noted that Fable's guard rails were so aggressive that they were the source of humor in the cyber community on launch day. Cyber security professionals found the model was blocking legitimate defensive research. The guardrails were if anything too strict. In her blog post, Missouri suggested that opponents of the expert control should print t-shirts with fix this code on the front and this shirt is ammunition on the back, which is a great way to guarantee you spend an extra 3 hours in the TSA security line getting intimately acquainted with airport security. The parallels between that era and this one are just not subtle at all. In the free fable open letter signed by over a 100 cyber security professionals from Nvidia, Adobe, Zoom, Google, and others made a further point. This capability is not unique to Fable. Open AI's GPT 5.5, Antropic's own other cloud models, and Chinese models including Moonshot AI's Kimmy 2.7 can all perform similar code reviews. The stated justification for pulling fable that it provides a unique uplift beyond other models does not hold against the evidence. The letter warns that pulling the best defensive tools from cyber security professionals while adversaries capabilities advance is dangerous.

Now, here is where the story stops being about a jailbreak and starts being about something else entirely. In February 2026, the Pentagon demanded that Anthropic allow its AI to be used for all lawful purposes, including fully autonomous weapons and mass domestic surveillance. Anthropic CEO Daria Amadei publicly refused. He said Anthropic cannot in good conscious exceed their request. He stated that autonomous weapons and mass surveillance are simply outside the bounds of what today's technology can safely and reliably do. The Pentagon's under secretary for research and engineering, Emil Michael, responded on X by calling Amadei a liar with a god complex who wants nothing more than to try to personally control the US military because nothing says mature national security discourse quite like highranking defense officials having an absolute public meltdown on X. On February 27th, President Trump ordered all federal agencies to immediately stop using anthropics technology. The Pentagon designated Antropic a supply chain risk, which is a classification normally reserved for companies with ties to foreign adversaries. And within days, Open AI announced a Pentagon deal. Sam Alman positioned Open AI as the cooperative alternative. Sam Alman basically sprinted to the Pentagon with a tray of cookies screaming, "I'll let you do whatever you want. Please pick me." He later described Antropic's handling of myths as fear-based marketing, saying, I'm quoting, "It is clearly incredible marketing to say, "We have built a bomb. We were about to drop it on your head. We will sell you a bomb shelter for $100 million." Which is ironic because Sam Alman's entire brand is literally, "We are building a sci-fi god that might destroy humanity. Please give us another $80 billion." And now 3 months later, that same administration has ordered Anthropics models pulled. The jailbreak was found by Amazon, which is simultaneously Anthropic's largest investor and a competitor, via its own bedrock and Titan AI platforms. With investors like these, who needs corporate espionage, right? Amazon literally snitched on its own portfolio company. The CEO, Andy Jasse, didn't even drop a polite Slack message. He went straight to the Oval Office. This is not just a conflict of interest at this point. That is successional levels of corporate betrayal. Amazon is out here playing 3D chess while funding the board, writing the rules, and calling the referee to report their own pieces for moving too fast. He personally called the White House about it. The directive arrived at 5:21 p.m. on a Friday. The time slot historically favored for announcements you'd rather people didn't pay too much attention to. And Axius reported that the administration was further inflamed because Entropic asked a cyber security expert to review the jailbreak whom they viewed as a radical Democrat and because Chris Krebs, the election security official Trump fired in 2020, vouched for her analysis. I want to be very careful about drawing conclusions here. I don't have inside knowledge of the government's reasoning, and I recognize that cyber security is a legitimate concern, but I can observe a pattern. A company refuses to allow its AI to be used for autonomous weapons. That company is banned from government contracts. That company's competitor gets the deal. That company then releases its most powerful model. That model is being pulled by the same administration triggered by research from the company's own investor competitor on a Friday evening. The timing is also worth noting from a commercial context. Anthropic had filed a confidential IPO perspectus worth a reported valuation of around $965 billion. Having your flagship model pulled by the government days before going public is not exactly helpful timing. That is a catastrophic vibe check for your valuation. Your IPO perspectives goes from we are reshaping the future of human intellect to please don't look at our Friday afternoon regulatory execution. Nothing says stable long-term investment quite like the Department of Defense treating your software update like a literal shipment of contraband. At this rate, Antropic's next funding round isn't going to be pitched in Silicon Valley boardrooms. It's going to be negotiated in some undisclosed underground bunker using coded knocks. Anthropic's own response was unusually direct for a company in its position: "We disagree that the finding of a narrow potential jailbreak should be the cost for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers." That is a company saying in public that the government's logic, if applied consistently, will shut down the entire AI industry. And they have a point for a country that treats free market principles as a founding value. This does not feel like a lot of free market. It feels like a sequence of events that rewards cooperation and punishes independence.

And on the autonomous weapons point, I want to state clearly where I personally stand. I agree fully with Daria Made on this one. No autonomous weapons, only God or circumstance should be making decisions about who lives or who dies. Not even us humans, allegedly the most developed species, should have the right to decide on the continuation of another creature's lives. This is not a technological limitation. It is a moral principle and it should not be controversial to say this. There is also an irony here that I think is worth sitting with because it has implications for every AI company that will ever make a safety disclosure. Anthropic was by most accounts the most transparent AI company about the dangers of its own technology. They said it. It told the world that Mythus could find vulnerabilities in every major OS and browser. It restricted the model rather than releasing it. It created Project Glasswing specifically to ensure only vetted organizations could access it. And then it built Fable with guardrail specifically designed to prevent the cyber security capabilities from being misused publicly. That transparency was used as the justification for shutting them down. As TechCrunch put it, the very caution Anthropic displayed in restricting mythos, has now apparently attracted exactly the kind of government scrutiny that could disrupt its business most. Sam Alman, who spent months calling this fear-based marketing, must be watching with considerable satisfaction. I bet he told the world Antropic was exaggerating. The government heard Antropic's own warnings and decided they were telling the truth. The responsible company got punished. The lesson basically for every other AI company is if you find something dangerous, be quiet about it. That lesson makes everyone less safe. All of us. And it is exactly the opposite of what good AI governance should incentivize. But I want to end this video with something that's been on my mind throughout all of this research because I think it matters a lot more than the politics. This channel has spent the past several weeks covering what I believe are the most important stories in AI right now. AI is measurably changing how human brains process information and the research suggests developing minds may never build cognitive abilities they outsource to machines. Cities are banning data centers because the communities paying for AI infrastructure are not the communities benefiting from it. 142,000 people lost their jobs in 5 months this year while the companies that fire them posted record revenue. Enterprise AI cost are spiraling and nobody has figured out how to make the economics work. Surveillance cameras are being covered with bin bags because cities cannot figure out how to turn them off. The most popular search engine on Earth was redesigned around AI that gets it wrong 57 million times per hour. But hey, at least it confidently tells you to add non-toxic glue to your pizza sauce to keep the cheese from sliding off. Those are the real fires. Those are the problems that affect millions of people right now in ways that shape their lives, their children's cognitive development, their communities, their employment, their privacy. And the US government is spending its political energy and capital on a three-word jailbreak that a 100 cyber security experts say is not unique, is not fixable, and whose removal actively harms defensive security. I am not saying jailbreaks do not matter. I am not saying cyber security is unimportant, but there is a skyscraper burning down and what appears to be happening is a detailed complaint about the local restaurant changing the menu. Yes, I know the building is collapsing, your honor, but the AI just gave somebody an unoptimized Python script, so we must freeze global commerce immediately. How will we ever cope? The priorities do not match the scale of the problems. And when you look at who benefits from those misplaced priorities, who gets the Pentagon contracts, who gets their competitor's model pulled, who gets to IPO without interference, the picture becomes harder to read as a straightforward national security decision and easier to read as something considerably more strategic. I think the people making these decisions should be spending their time on the things that will actually determine whether AI makes the world better or worse for the billions of people now living with it. the cognitive impacts, the surveillance infrastructure, the job displacement, the enterprise cost crisis, the data center backlash, the communities being affected. We have no shortage of nonsense happening to humanity right now. Those are the issues where political energy and regulatory attention could make a genuine difference. A three-word jailbreak on a model that a 100 cyber security experts say is no more capable than its competitors doesn't really feel like an emergency to me. The emergency is everything else, including the climate crisis, by the way. I'm not minimizing that. And the longer the attention stays in the wrong place, the more time the actual fires have to spread. But Anthropic's troubles are only one part of a much broader picture. Open AAI, the company that got the Pentagon deal, the company whose CEO mocked Antropics Transparency, is simultaneously losing $1.22 for every dollar it makes and preparing a trillion dollar IPO. Losing money on every single transaction while expecting a trillion dollar valuation is the ultimate techra magic trick. Truly, the math is mathing. I cover that in detail and it connects directly to everything happening here. That's the video that I would watch next. Thank you so much for watching this one.