Updated June 12, 2026.
I have been writing on this blog for years about artificial intelligence, security, and the way these technologies intersect with politics and our daily lives. Rarely, though, have I come across a story as revealing as today’s. Anthropic, one of the leading companies in language models, has had to abruptly shut down two of its models —Fable 5 and Mythos 5— because the United States government ordered it to. I want to walk you calmly through what happened, why it matters, and what I think all of this says about the moment we are living in.
What happened
The US government, citing its national security authorities, has issued an export control directive suspending all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States. This includes, oddly enough, Anthropic’s own foreign national employees. The practical effect is devastating: in order to comply, the company has had to abruptly disable both models for all of its customers. The rest of Anthropic’s models, they explain, are not affected.
According to the company’s own statement, they received the directive that same day at 5:21pm (ET). The letter did not specify exactly what the national security concern was. Anthropic’s understanding is that the government believes it has discovered a method for bypassing the model’s safeguards —what is known in the field as a jailbreak.
The heart of the matter: the “jailbreak”
This is where things get interesting, and where I want to pause. A jailbreak is, put simply, a technique to trick an AI model into doing something its safeguards are supposed to prevent. Anthropic says it reviewed a demonstration of the specific technique that alarmed the government, and its conclusion is that it served to identify a handful of minor, already known vulnerabilities. What’s more, they argue that those same vulnerabilities can be found by other publicly available models without needing any trick at all.
The company recalls that it locked Fable down with very strict safeguards —so strict, they say, that many users complained they were excessive— and that they spent thousands of hours, together with the US government, the UK AI safety institute, and several external organizations, red-teaming the model. Their thesis is that no one has yet achieved a universal jailbreak, that is, a method capable of broadly disabling the model’s protections. They do acknowledge, however, that non-universal jailbreaks exist, which under very specific circumstances may extract some information, and they honestly admit that perfect resistance to these attacks is probably not possible today for any provider.
The most striking detail is what the alleged jailbreak behind the order actually consists of, according to Anthropic: essentially, asking the model to read a specific source code and fix any software flaws it finds. In other words, something any developer does every day and that other models, such as OpenAI’s GPT-5.5, also offer without much trouble.
Why this matters
Anthropic says it is complying with the order and removing access, but makes its disagreement clear: it does not consider it reasonable that the discovery of a narrow, non-universal jailbreak should justify pulling a commercial model used by hundreds of millions of people. And it warns of something I share: if that standard were applied across the entire industry, it would essentially halt the launch of any new model.
To me, what is relevant about this case is not just the technical detail, but the precedent. We are watching, in real time, how a government decides to shut down a massively used tool based —according to the company itself— on verbal evidence and a concern that has not been fully detailed. It is the tension between state control and technology deployment, and it is a debate that will be with us for years.
What comes next
I don’t want to close this article without telling you what interests me most about all of this. In upcoming posts I will spend time talking specifically about prompt jailbreaking: what it is, how it works, why it is so hard to prevent, and what implications it has for the security of the models we already use every day. I believe that understanding this phenomenon well is key to making sense of stories like today’s, so stay tuned.