The US ordered Anthropic to kill its two most powerful AI models — overnight

Information Technology 3 min 80 sources

A Commerce Department directive forced Anthropic to shut off Fable 5 and Mythos 5 for every user worldwide, days after launch. The trigger: the company's own warnings about how dangerous the models were.

Key takeaways

The US Commerce Department ordered Anthropic to shut off its two most powerful AI models on Friday night; the company complied worldwide within hours but says the government got it wrong.
The trigger was a disputed claim that someone jailbroke Fable 5 to find software flaws — a capability Anthropic says is minor and already common in rival models.
A benchmark-leading model went dark globally on a single directive with no warning — a reminder of how fragile it is to build on one vendor's top model.

The US government ordered Anthropic on Friday to immediately shut off access to its two most capable AI models, Claude Fable 5 and Claude Mythos 5. Anthropic complied that night — and made clear it thinks the government got it wrong [1][4].

The Commerce Department directive arrived at 5:21 pm ET Friday, framed as an export control: it restricts the models to inside the United States and bars foreign nationals [4][30]. But to comply in the immediate term, Anthropic said the only option was to disable both models for all customers, everywhere — not just the foreign users the order nominally targeted [1][30]. Access to Anthropic’s other models is unaffected.

Why these two models

Mythos was Anthropic’s most powerful model — previewed in April, then kept tightly locked away ever since [4]. The reason the company gave: it was unusually good at finding security holes in software. Anthropic said Mythos identified flaws in every major operating system and web browser it tested. Rather than release it, the company ran a controlled program — Project Glasswing — sharing it with roughly 50 vetted organisations including Amazon, Apple, Google, Microsoft and CrowdStrike, for defensive cybersecurity work [4].

Fable 5 was the commercial version: Mythos fitted with guardrails — automated filters that block answers in high-risk areas like cybersecurity and biology — to make it safe enough for general release. Anthropic launched it just three days ago. By one independent benchmark from Vals AI, a firm that tracks model performance, it was immediately the most capable AI model available to the public [4].

The trigger: a claimed jailbreak

A “jailbreak” is a prompt that tricks a model into ignoring its own safety rules. According to an Axios report cited by Ars Technica, the administration acted on reports that someone had found a way around Fable 5’s filters — specifically, to get it to review a codebase and flag software flaws [30].

Anthropic’s account is narrower. It says the government has shown it only “verbal evidence of a potential narrow, non-universal jailbreak” — one that surfaced “minor” and “relatively simple” vulnerabilities, and a capability already present in other public models like OpenAI’s GPT-5.5 [4][30]. The company also says its strongest safeguards run as separate filter systems outside the model, so getting Fable to keep talking past a refusal doesn’t disable them — and a review of recent use found no case where harmful content actually got through [4].

The administration reportedly asked for a pause to give the “national security apparatus” a few weeks to harden against the threat [30]. Anthropic called the disruption a “misunderstanding” and promised more details within 24 hours [30].

The part worth sitting with

Anthropic spent months telling the world how dangerous Mythos was — that it was too capable to release, that it could break every browser it touched. That story was the company’s safety credential. It was also, it turns out, a map of exactly where to grab.

“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 wrote. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers” [4][30].

For anyone building or depending on frontier AI, the practical signal is the speed and reach: a single directive, received at 5:21 pm, took a benchmark-leading public model offline by nightfall — worldwide, on a claim the vendor disputes and hasn’t seen written down. If your product, your security workflow, or your codebase runs on one company’s top model, that model can vanish on a Friday evening with no notice. Earlier this month President Trump signed an executive order urging AI makers to submit to voluntary government security testing [30]. “Voluntary” now sits next to “and we can switch you off.”

02 · Lesson · why it matters

The reputation you build is a map someone else can read

The story you tell to be trusted also tells everyone exactly where your power sits — and what it's worth seizing.

A safety credential, used as a target

For two months, Anthropic told the world a careful story about itself. Its top model, Mythos, was so good at finding holes in software that releasing it would be reckless. So the company locked it away and handed it only to fifty trusted organisations for defence. The message was: we are the responsible ones. Our model is too powerful to set loose.

That story was the company’s reputation. It was also a set of directions.

On Friday at 5:21 pm, the US government read those directions back to Anthropic and ordered both models switched off — worldwide, by nightfall. The official reason was a disputed jailbreak. But the deeper reason was the story Anthropic had spent months telling: this thing is dangerous. You cannot build your name on “our model could break every browser” and then be surprised when someone with authority over danger decides they should hold the switch.

A reputation isn’t private

We tend to think of a reputation as something we own — a quiet asset, ours to spend. It isn’t. A reputation is a public coordinate. It tells other people where you are, what you’re good for, and what you’re worth taking.

The careful version of this is everywhere. A company that markets itself as “the secure one” invites the regulators who police security. A worker who becomes known as the only person who understands the old system becomes the person no one will let leave. A country that advertises a weapon as a deterrent has also told its rivals exactly which capability to match. The signal you send to earn trust is the same signal that draws the people who want to control, copy, or constrain you.

Anthropic’s whole identity was being the safety-conscious alternative — the grown-up in the room. That identity is real, and it’s why many people trust the company. It was also the precise feature that made it the first one ordered to shut down. The story did its job too well.

The thing you don’t hold

Here is where it stops being only about one company. When Mythos and Fable went dark, they didn’t go dark just for Anthropic. They went dark for everyone downstream.

Fifty organisations — Amazon, Apple, Google, Microsoft, CrowdStrike — were using Mythos to find security flaws before attackers did. That work paused on Friday night. Developers who had wired Fable 5 into their products three days earlier woke up to a model that no longer answered. None of them did anything wrong. None of them were even the target — the order named foreign nationals, and it took down the model for all of them anyway.

They were inside the system without holding any of its controls. The switch sat in two places: a government office that sent a directive, and a company that complied within hours. Everyone else was a node — connected to the outcome, absent from the decision. If your product runs on one vendor’s best model, you are one of those nodes too. You inherit every story that vendor tells about itself, including the ones that get them noticed.

What a single switch reveals

The speed is the lesson inside the lesson. One directive, one evening, and a benchmark-leading model used by hundreds of millions was gone — on a claim the vendor says it has only heard out loud, never seen written down. No appeal that mattered in the moment. No notice.

That is what it looks like when a great deal of capability, and a great deal of dependence, runs through a single point that one actor can close. It isn’t a flaw in this company or that government. It’s the shape of building on something you don’t control, dressed up by a reputation you thought was protecting you.

Holding it loosely

So the careful question isn’t “was the government right or wrong” — the facts there are still disputed and incomplete, and neither of us is in the room where the jailbreak was or wasn’t demonstrated. The more useful thing to carry is smaller and harder to see: the very strengths we advertise to be trusted are the same coordinates that make us reachable.

Anthropic told a true story about being careful, and the story became a lever. The fifty defenders and the early adopters told no story at all, and still lost their model. From inside any single seat — the company’s, the regulator’s, the developer’s — the whole picture is invisible. Each one sees a reason, a rule, a disruption. None of them sees the full web that turned a safety credential into a kill order overnight. That web is the thing worth respecting, precisely because no one standing in it can see all of it.

03 · Lab · your turn

One Switch You Don't Hold

Distribute your product's critical parts across AI vendors, then a directive takes one offline — and you feel what concentration costs.