The very caution that proved responsible now attracts the scrutiny that threatens it.
In a moment that reveals the deepening tension between state power and the frontier of artificial intelligence, the U.S. government ordered Anthropic to shut down its two most capable AI models globally, citing national security concerns over a discovered jailbreak. Anthropic complied, but publicly contested the justification, arguing the vulnerability is narrow, already present in a competitor's widely available model, and insufficient grounds for so sweeping an action. The episode raises a question that will outlast this particular dispute: whether the companies most willing to be transparent about risk are the ones most likely to be punished for it.
- With a Friday evening directive, the U.S. government invoked export controls to force an immediate, worldwide shutdown of Claude Mythos 5 and Fable 5 — two of the most powerful AI models ever made available.
- The trigger was a claimed jailbreak of Fable 5, but Anthropic pushed back hard, calling it a narrow, non-universal vulnerability no more dangerous than what already exists inside OpenAI's publicly available GPT-5.5.
- Anthropic's safety architecture — independent classifier networks designed to hold even when a model is manipulated into responding — was dismissed or ignored, leaving the company arguing into a void.
- The shutdown lands at the worst possible moment: Fable 5 had been live for just three days and was already benchmarked as the most capable public AI model, while Anthropic is widely expected to pursue an IPO this year.
- The deepest wound may be philosophical — Anthropic's years of public safety-first positioning, the very thing that made it legible and trustworthy to regulators, appears to have made it the most visible target.
Late on a Friday, Anthropic received a government directive that would go into effect within hours: shut down Claude Mythos 5 and Claude Fable 5, everywhere, for every user on Earth. The company complied — but it did not go quietly, publishing a detailed public statement that amounted to an open dispute with the government over whether the action was warranted.
The official framing was export controls, but Anthropic understood the real concern to be a jailbreak discovered in Fable 5 — specifically, a method of prompting the model to analyze code and identify security vulnerabilities. The company's response was pointed: this is not exotic capability. The same functionality exists in OpenAI's GPT-5.5, available to the public, and is used routinely by cybersecurity professionals for defensive work. A narrow, non-universal vulnerability, Anthropic argued, should not trigger a global recall — and if that standard were applied consistently across the industry, no frontier model could ever be deployed.
To appreciate the stakes, consider what these models are. Mythos is Anthropic's most powerful system, so capable at finding software flaws that the company chose not to release it publicly at all, instead sharing it with roughly fifty vetted organizations — Amazon, Apple, Google, Microsoft — through a controlled program called Project Glasswing. Fable 5 was the commercial answer: Mythos fitted with safety guardrails, released just three days before the shutdown, and immediately ranked by independent benchmarks as the most capable AI model available to the general public.
The shutdown stands, and the irony is difficult to miss. Anthropic has staked its identity on being the safety-conscious voice in a reckless industry — restricting its most powerful model, warning publicly about frontier AI risks, building its brand around restraint. That same restraint made the company legible to regulators, and now it has made the company a target. With an IPO widely anticipated this year, a global model shutdown is precisely the kind of headline Anthropic could not afford — and the kind its competitors, quieter about their own risks, have so far avoided.
On Friday evening at 5:21 pm Eastern time, Anthropic received a directive from the U.S. government that would force the company to pull the plug on two of its most capable artificial intelligence models from every user on Earth. Within hours, Claude Mythos 5 and Claude Fable 5 went dark globally. Anthropic complied with the order but made its disagreement unmistakable, publishing a lengthy statement that amounted to a public argument with the government over whether the action was justified.
The government framed the shutdown as an export control measure, nominally designed to keep the models away from foreign nationals. But Anthropic's reading of the underlying concern was more specific: the government believed someone had found a way to jailbreak Fable 5, bypassing its safety guardrails. The company's frustration was evident in how it described what the government had actually shown them—a "potential narrow, non-universal jailbreak" that essentially amounted to prompting the model to analyze a specific piece of code and identify security flaws within it. This, Anthropic pointed out, is not some exotic capability. The same level of functionality already exists in OpenAI's GPT-5.5, available to the public. Cybersecurity professionals use this kind of analysis routinely as part of their defensive work.
To understand why this moment matters, you need to know what these two models represent. Mythos is Anthropic's most powerful creation, a system so capable at finding software vulnerabilities that the company decided it was too dangerous to release broadly. When Anthropic tested Mythos against every major operating system and web browser, it found flaws in all of them. Rather than make it public, the company created Project Glasswing, a controlled program that shared Mythos with roughly fifty vetted organizations—Amazon, Apple, Google, Microsoft, CrowdStrike—for defensive cybersecurity research. Fable 5 was Anthropic's compromise with commercial reality. Released just three days before the shutdown order, it was essentially Mythos fitted with safety guardrails designed to block responses in high-risk domains like cybersecurity and biology. According to performance benchmarks from Vals AI, Fable 5 was immediately the most capable AI model available to the general public.
Anthropic's argument against the shutdown rested on two pillars. First, the company maintained that its safety systems operate through independent classifier networks that function separately from the model itself. Even if someone managed to convince Fable 5 to keep responding past a refusal, the underlying protections against the most dangerous outputs would theoretically remain intact. Second, and more broadly, Anthropic contended that the government's standard for action was dangerously low. "If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers," the company wrote. A narrow, non-universal jailbreak discovered in a commercial model deployed to hundreds of millions of people should not, in Anthropic's view, trigger a global recall.
None of this persuaded the government to reconsider. The shutdown stands. And the irony cuts deep for Anthropic, which has built its public identity around being the safety-conscious alternative to competitors like OpenAI. The company's caution—its decision to restrict Mythos because it was too powerful, its public warnings about the risks of frontier AI—appears to have backfired. The very restraint that Anthropic promoted as evidence of responsible stewardship has now attracted the kind of government scrutiny that threatens to disrupt the company's business at a critical moment. Anthropic is widely expected to pursue an initial public offering this year. A global shutdown of its two most powerful models, even if temporary, is not the kind of headline a company wants in the lead-up to that event.
Citas Notables
We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.— Anthropic, in blog post responding to government shutdown order
If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.— Anthropic, arguing the government's threshold for action is too low
La Conversación del Hearth Otra perspectiva de la historia
Why would the government shut down models that Anthropic itself restricted? Doesn't that suggest Anthropic was already being cautious?
Exactly. Anthropic's caution is what drew the attention. By publicly saying Mythos was too dangerous to release, they essentially told the government, "This thing is powerful enough to be a national security concern." The government took them at their word.
But Anthropic says the jailbreak the government found isn't even unique. It exists in OpenAI's models too.
That's their argument, yes. And it's a fair one—if you're going to shut down Fable for a narrow vulnerability, why not shut down GPT-5.5? But the government may have been responding to Anthropic's own safety warnings, not conducting a level playing field audit.
So Anthropic's transparency about risks backfired?
In a way. They were trying to be responsible by restricting Mythos and explaining why. But that transparency gave the government a roadmap for where to look and what to worry about.
What happens to Anthropic now?
They're planning an IPO this year. A global shutdown of your two most powerful models weeks before going public is a serious problem. Even if it's temporary, it's a signal that regulators see these systems as risky.
Could they have avoided this by being less transparent about the risks?
Possibly. But that's a grim calculus—stay quiet about dangers to protect your business, or be honest and risk government intervention. There's no good answer there.