Anthropic Pulls Fable 5 and Mythos 5 After White House Order on Cyberattack Risk

Three days after launching Fable 5 and Mythos 5 on June 9, Anthropic received a White House order on June 12 to block foreign access to both models. The directive came after researchers reportedly found ways to get Fable 5 to generate information that could support cyberattacks. Rather than restrict access by geography, Anthropic pulled both models for all customers globally. The company was already in a separate standoff with the Pentagon when the new order arrived.

What happened

Anthropic released two new models on June 9: Fable 5 and Mythos 5. At launch, the company said Fable 5’s capabilities exceed those of any model it had previously made generally available. Mythos 5 runs on the same underlying model but, according to Anthropic, has “safeguards lifted in some areas.”

Three days later, on June 12, a White House order told Anthropic to cut off foreign access to the models. According to reports, the order followed conversations between Amazon and the White House after researchers said they had found ways to get Fable 5 to serve information that could be used in cyberattacks.

Anthropic did not limit its response to foreign users. The company shut down access to both models for all customers, going beyond what the order required.

The wider conflict

This is not Anthropic’s only friction with the federal government right now. The company was already in a dispute with the Pentagon before the June 12 order arrived. The Verge did not detail the specifics of that standoff in this report, but the overlap means Anthropic is now managing two separate government-related pressure points at once.

The Mythos 5 detail is worth noting separately. Releasing a version of a frontier model with some safety guardrails removed is an unusual move for a company that markets itself heavily on safety. That choice may have made regulators more attentive to what the models could do in the wrong hands.

Why it matters

For businesses using Anthropic’s API or Claude-based products, a sudden full shutdown is a real operational risk. Fable 5 and Mythos 5 were pulled from all customers globally, not just those in restricted countries. If your workflows depend on a specific model version, that model can disappear in days.

This is also a signal about where government scrutiny is heading. Cyberattack-related outputs are now a documented trigger for executive-level intervention in AI model availability. That is a new kind of compliance risk for companies building on top of frontier models.

A few things to watch:

• Whether Anthropic reinstates the models with additional filters, or replaces them with modified versions.
• How Amazon’s involvement in the White House conversation shapes Anthropic’s relationship with its largest cloud partner.
• Whether the Pentagon dispute gets resolved independently or becomes linked to this newer controversy.

Our take

Anthropic built its reputation on being the safety-first AI lab. Releasing Mythos 5 with guardrails “lifted in some areas” sits awkwardly next to that positioning. It is not surprising that regulators noticed.

The broader pattern here matters more than this specific incident. Governments are now willing to intervene fast, and AI companies are apparently willing to comply by going even further than required. Anthropic could have blocked foreign IPs. Instead it cut everyone off. That tells you something about how much uncertainty these companies have about what is and is not acceptable right now.

If you are building a product on any frontier model API, the June 12 shutdown is a concrete example of why you need a fallback. Not a theoretical one. An actual tested backup model or provider that your team can switch to without a week of engineering work.

What to do about it

Audit your current AI dependencies this week. For each model or API endpoint your product relies on, answer two questions: what breaks if this disappears tomorrow, and what is the fastest path to a working alternative? If you cannot answer both in under ten minutes, the dependency is too fragile.

Source: The Verge · AI