Anthropic Disables Fable 5 and Mythos 5 on US…

On June 12, 2026, at 5:21 p.m. ET, Anthropic received a US government directive ordering the immediate suspension of all access to Fable 5 and Mythos 5 by foreign nationals, including the company's own foreign employees. Anthropic disabled both models for all global customers. The government cited an alleged jailbreak method it believes threatens national security; Anthropic contends this is a distorted reading of unverified verbal evidence.

On June 12, 2026, at 5:21 p.m. ET, Anthropic received a US government directive ordering the immediate suspension of all access to Fable 5 and Mythos 5 by foreign nationals, including the company's own foreign employees. Anthropic disabled both models for all global customers. The official rationale: an alleged jailbreak method the government believes threatens national security. Anthropic maintains this is a distorted reading of unverified verbal evidence.

Key Takeaways

The directive, received June 12 at 5:21 p.m. ET, compels Anthropic to bar access to Fable 5 and Mythos 5 for every foreign national, inside and outside the United States
Anthropic deactivated both models for all global customers because it could not guarantee nationality-based filtering without disrupting service for everyone
The government provided only verbal evidence of a "narrow and non-universal" jailbreak, consisting of asking the model to analyze a codebase and fix software flaws
Anthropic warns that applying this standard across the industry would halt every new frontier model deployment
Access to other Anthropic models, including Claude Opus 4.8, was unaffected

A Block Without Technical Advisory: What Exactly the Government Demands

The directive has not been made public. CyberScoop identifies Commerce Secretary Howard Lutnick as the signatory; other sources describe it generically as a federal government order. What is known comes almost entirely from Anthropic's statement, issued via a press release and reported by specialized outlets.

According to Anthropic, the government believes it discovered a method to bypass Fable 5's protections. The technique, described by the company as "essentially consisting of asking the model to read a specific codebase and fix any software vulnerabilities," was not documented in writing. Anthropic says it reviewed the demonstration and found minor, already-known vulnerabilities reproducible with other publicly available models, including OpenAI's GPT-5.5.

The restriction hits not only foreign users but also Anthropic's foreign employees on US soil. Traditional export regulations govern the transfer of technology across borders, not access by foreign workers on American territory. The government's text has not been released, nor is it known whether exemptions exist for residents with work permits.

"The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance."
— Anthropic, official statement

The Technical Dispute: Real Jailbreak or Overinterpretation?

The core of the dispute is the nature of the threat. According to Anthropic, no universal jailbreak was found by its own red teams after more than a thousand hours of external testing. The fallback rate for Fable 5's safeguard triggers is below 5%, according to pre-directive data gathered by The Hacker News. The model was designed with a system of overlapping classifiers to make simultaneous removal of all constraints difficult.

Anthropic contests that the government interpreted a model capability excessively. Automated codebase analysis to discover vulnerabilities is exactly what the Glasswing project, Mythos 5's partner, demonstrated: over ten thousand high- or critical-severity flaws found, of which 1,726 were confirmed true positives, with 1,094 rated high or critical severity. These numbers, documented before the directive, show the model was employed as a vulnerability discovery tool.

The government, however, appears to have received the information through a different channel. David Sacks, a government official with an AI portfolio, wrote on X that Anthropic "refused to fix a jailbreak" reported by a "highly credible trusted partner." The identity of this partner is unknown. Multiple sources — The Information, Reuters, Wall Street Journal per The Hacker News — point to Amazon and CEO Andy Jassy as the parties that raised the alarm, but Amazon has neither confirmed nor denied. Semafor added an unverified element: concern that a China-linked group had access to the model. It is not documented how the government learned of this possibility nor whether it was substantiated.

Institutional Reaction: Military Support and Process Criticism

The directive drew public support from DoD CIO Kirsten Davies, who wrote on X: "Some things are simply more important than revenue cycles, clickbait, and pre-IPO valuation. America First. Always." The message, reported by CyberScoop and NextGov/FCW, explicitly links the decision to the ongoing dispute between Anthropic and the Pentagon.

That dispute has prior roots. On March 27, 2026, a federal judge issued a temporary injunction against the Pentagon's designation of Anthropic as a "supply chain risk," decided in February after negotiations on a military contract collapsed. The June 12 directive arrives ten days after a presidential executive order on AI that established a new model verification mechanism. The temporal overlap fuels suspicions, reported in sources, that the export control directive may be instrumental to the contract dispute.

Across the Atlantic, UK Minister for AI and Online Safety Kanishka Narayan raised concerns about technological sovereignty. Dean Ball of the Foundation for American Innovation called the move "simply cartoonish," in a comment reported by AI Business.

Commercial Context: IPO at Risk and Regulatory Precedent

Anthropic had filed for its IPO in early June 2026, with an expected valuation exceeding one trillion dollars according to the Financial Times. The directive arrives two days after CEO Dario Amodei published a policy essay asking the government for legal authority to block deployment of unsafe AI.

The June 12 directive is not the first government intervention on AI models. Ten days earlier, a presidential executive order had already established a verification mechanism for frontier models. This precedent makes any claim presenting the June 12 action as the first AI regulation inaccurate.

"If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers."
— Anthropic, official statement

What Changes

For organizations using Fable 5 or Mythos 5, the interruption is total and immediate. Anthropic has not indicated timelines for reactivation. Access to Claude Opus 4.8 and the company's other models remains active.

The dispute raises questions about the standard of proof required to block a commercial model. Anthropic underscores that the same codebase analysis capability exists in OpenAI's GPT-5.5, without triggering analogous directives. The discretion with which the government can act on verbal, unwritten evidence represents an element of uncertainty for all frontier model providers.

The case also highlights the risk of overlap between commercial disputes and regulatory tools. The Pentagon's designation of Anthropic as a supply chain risk, now suspended by a judicial injunction, and the export control directive emerge from the same institutional context. According to sources, suspicions exist that the two actions are linked, but this reconstruction has not been independently verified.

Limits of the Reconstruction

This reconstruction relies on Anthropic's statement and reporting by specialized outlets. The text of the government directive has not been made public. It was not possible to independently verify claims about Amazon, any China-linked groups, or the exact nature of the reported jailbreak technique. The government's position is not available directly: the Department of Commerce declined to comment.

Information has been verified against cited sources and is current as of publication.

Sources

Sources and references