Did Anthropic Do This to Itself? Inside the Fable 5 and Mythos 5 Shutdown

Anthropic spent years warning the world that frontier AI could become dangerous enough to demand government oversight. Days after shipping the most capable models it had ever built, the U.S. government appears to have taken the company at its word — and Anthropic is now living with the consequences.

Within days of release, Anthropic disabled both Claude Fable 5 and Claude Mythos 5 worldwide, citing a directive from the Trump administration. Fable 5 had launched as the public-facing edge of Anthropic's technology; Mythos 5 was a far more capable system offered to a narrow set of users. Then both were gone — not throttled, not regionally restricted, but switched off globally. It is, by any measure, one of the most consequential AI-policy events to date: the first time the U.S. government has directly forced a frontier model family offline over national-security concerns.

The most provocative reading of what happened comes from AI commentator Matthew Berman, whose video Fable and Mythos taken down by Trump frames the shutdown around an uncomfortable question — one worth sitting with: did Anthropic, the industry's loudest safety advocate, help build the case against itself?

What actually happened

The official trigger, according to reporting Berman cites, was a fear that the models could be jailbroken in ways that exposed dangerous cybersecurity capabilities. Government agencies reviewed the concern and escalated it to senior officials. What followed was extraordinary in its speed: Anthropic was reportedly given roughly 90 minutes to act before export-control measures kicked in.

Those controls would have barred access by foreign nationals — including, awkwardly, some of Anthropic's own employees. With no practical way to enforce that restriction selectively and instantly, the company reached for the only lever it fully controlled: it took the models down everywhere. A global shutdown wasn't a statement of agreement. It was the path of least resistance under an impossible deadline.

The Amazon angle

A central thread in Berman's account is the alleged role of Amazon — one of Anthropic's largest investors and a key cloud partner. By this telling, Amazon flagged the issue to U.S. officials after researchers discovered methods for bypassing some of the models' safeguards, and that warning is what set the review in motion.

Anthropic's counter-argument is that the vulnerabilities were narrow, limited in scope, and not meaningfully different from weaknesses found across other leading AI systems. In other words: real, but not uniquely catastrophic — and arguably not the kind of thing that warrants pulling an entire model family from the global market in an afternoon.

Advertisement

Serverless Inference by DigitalOcean 55+ models, every modality. One API key, one bill.

Did Anthropic do this to itself?

This is where Berman's framing gets pointed, and where the story stops being a simple government-versus-company dispute.

Anthropic has built its identity on the premise that frontier AI is powerful enough to be hazardous and therefore needs guardrails, evaluations, and yes — government attention. That message has been a genuine differentiator. But messaging has gravity. If you spend years telling regulators that the most advanced models could enable serious cyber or bio harm, you should not be shocked when regulators eventually treat one of your most advanced models as exactly that: a strategic asset to be controlled, not a product to be sold.

Several commentators in the surrounding coverage make a sharper version of the point: the company's own rhetoric about the dangers of frontier AI may have lowered the bar for officials to justify treating its models as national-security matters. The very vocabulary Anthropic helped popularize — "frontier," "catastrophic risk," "dual-use" — is now being used to take its products away.

None of this means Anthropic was wrong to care about safety, and Berman's argument is a theory, not a verdict. The vulnerabilities may have been more serious than the company admits, or the administration's response may have been a dramatic overreach. But the irony is hard to miss: the safety-first lab got the safety-first outcome it had been describing all along — just not in the form it wanted.

Why this is bigger than one company

Strip away the specifics and a precedent remains. Until now, "AI safety meets national security" has been mostly a conference-panel abstraction. Here, policymakers moved fast and decisively to pull a specific model family offline. That is a template other governments can copy.

If frontier models start being governed like advanced semiconductors — subject to export controls, access gated by nationality and jurisdiction, deployment contingent on government sign-off — the implications cascade quickly:

• Access becomes political. Who can use the best models may depend less on willingness to pay and more on where you live and which passport you hold.
• Open-source gains urgency. Every hard restriction on closed frontier models is an argument for open-weight alternatives that no single government can switch off in 90 minutes.
• Other nations accelerate. Countries watching a U.S. lab get its flagship yanked on short notice have fresh incentive to build domestic capabilities rather than depend on American providers.

The takeaway

The shutdown of Fable 5 and Mythos 5 is less a one-off scandal than a glimpse of the operating environment ahead. Whether the government's concerns were justified or overblown, the episode shows that frontier AI has crossed into a new phase, where technical capability, national security, geopolitics, and regulation are braided together.

The lesson for AI companies is blunt: you are no longer operating only in a technology market. You are operating in a policy and national-security environment — and as Anthropic just learned, the stories you tell about your own power can come back to govern you.