How Anthropic's Safety Warnings Set a Trap the Government Was Happy to Spring

On June 9, 2026, Anthropic launched Claude Fable 5, describing it as state-of-the-art across nearly all tested benchmarks and adding that its cybersecurity capabilities warranted safety guardrails no other AI company had bothered to implement. Three days later, at 5:21 p.m. ET on June 12, the US Commerce Department issued an order requiring Anthropic to shut down Fable 5 and Mythos 5 for all users worldwide. The reason cited: national security. The sequence was not random.

A Year of Safety Marketing

To understand the export order, it helps to recall what Anthropic spent the past year saying about Mythos. In April, when the company previewed the model through Project Glasswing, it published benchmark data showing Mythos could find software vulnerabilities at a scale no human team could match. The company cited these results not to sell the model, but to explain why it was not selling it. "Over 99% of the vulnerabilities we've found have not yet been patched," Anthropic wrote at the time. Releasing Mythos broadly, it argued, "would create an asymmetric risk: any attacker with API access could replicate the same discovery process, but defenders would not have time to ship fixes."

That framing continued through the Fable 5 launch. Fable 5 is a public version of Mythos with safety guardrails, and in describing why those guardrails were necessary, Anthropic was restating the case for why the underlying model was dangerous. The launch announcement noted that in high-risk domains including cybersecurity, Fable 5 would fall back to Claude Opus 4.8 rather than answer fully. The implication: the underlying capability is real enough to block.

Taking Anthropic at Its Word

Sam Altman flagged the dynamic in April. "It is clearly incredible marketing to say, 'We have built a bomb and we will sell you a bomb shelter for $100 million,'" the OpenAI CEO said during an industry panel. His complaint was that Anthropic was using safety language to charge premium prices and differentiate from competitors, not that the safety concerns were wrong. What he did not say was that the same language could eventually reach a government that would act on it literally.

When the Commerce Department received word that someone had found a method to bypass Fable 5's safeguards, it had months of Anthropic's own documentation to draw on. The company had described Mythos's cybersecurity capabilities in granular detail. It had explained why those capabilities warranted restricted deployment. It had published figures showing Mythos could autonomously chain vulnerabilities into complete exploit paths on hardened systems. The government's letter cited national security. That framing maps directly onto what Anthropic had been saying.

"If you describe your product as a munition in every press release, eventually a government takes you at your word." Peter Girnus, cybersecurity researcher, on X, June 12, 2026

Arguing the Opposite

Anthropic's position since receiving the export order has been almost the inverse of its public communications over the past year. The company now says the jailbreak is "narrow and non-universal," applying to a specific code-review scenario rather than unlocking the model's full capabilities. It argues that applying this standard across the industry "would essentially halt all new model deployments for all frontier model providers." Both points may well be correct. But they sit uneasily alongside a year of communications designed to establish that this particular model is categorically more dangerous than anything else on the market.

The timing compounds the awkwardness. On June 10, two days before the export order arrived, Dario Amodei published a sweeping policy essay calling for FAA-style mandatory testing of frontier AI models, with government authority to block releases that fail third-party evaluation. The proposed testing domains included cybersecurity, biological weapons, and the capacity of a model to accelerate autonomous research. Two days later, the government exercised precisely that kind of authority, citing a cybersecurity concern, on Anthropic. The company that argued most publicly for a more interventionist AI regulatory posture may have made it easier, not harder, for regulators to act.

Anthropic says it is working to restore access as soon as possible. The path back requires convincing the same government the company spent a year educating about Mythos's dangers that the risk, in this specific case, was overstated. That is not an impossible argument. But the export order is currently live, and there is no indication of how quickly the Commerce Department intends to revisit it. The lesson is simpler than the situation: when you spend a year telling the world your product is a weapon, some readers will believe you.