US Influences Anthropic’s Top AI Models

Secretary of Commerce Howard Lutnick sent a letter to Chief Executive Officer Dario Amodei on Friday, announcing that Anthropic has expanded coverage to foreign nationals both inside and outside the United States, including its employees. Because Anthropic was unable to verify the nationality of all users in real time, they disabled both models for all customers to ensure compliance.

Anthropic said the letter does not specify any underlying national security concerns and only provides verbal evidence of a possible Fable 5 escape.

Anthropic launched Fable 5 last week as the first public model of its new Mythos class, its most capable tier to date. Mythos 5 is the same basic model with certain safeguards removed, but it was still limited to Project Glasswing, Anthropic’s program for vetted cyber defenders and infrastructure providers.

Mozilla recently reported that it had solved hundreds of vulnerabilities using early Mythos models. Both models are priced at $10 per million input tokens and $50 per million output tokens (see below). Claude Mythos 5 can build exploits, but can’t drive campaigns).

Fable 5 was designed to make Mythos class functionality safe for general release. It uses a layer of classifiers, a separate AI system that detects potential misuse before the main model responds. If the query involves cybersecurity, biology and chemistry, or model distillation, Fable 5 will route the query to Claude Opus 4.8 instead and notify the user. Over 95% of Fable sessions do not trigger fallback.

Before launch, Anthropic said it spent thousands of hours stress testing Fable’s safety equipment in collaboration with the U.S. government, the UK AI Safety Institute, private third-party organizations, and internal teams. None of the testers found a universal jailbreak that could broadly bypass the model’s protections across a wide range of tasks.

The company said it created the cybersecurity classifier to thwart both the development of specific exploits and large-scale agent-like hacking activities such as reconnaissance, lateral movement, and multi-stage cyber operations. In testing, Fable 5 blocked progress on all of these tasks.

Anthropic has also introduced a 30-day data retention requirement for all Fable 5 and Mythos 5 traffic to support ongoing jailbreak research and mitigation efforts. The company said the data will not be used to train models and will be deleted after 30 days.

Anthropic reviewed the report underlying the government directive and said the capabilities demonstrated there are also available from other publicly deployed models, such as OpenAI’s GPT-5.5, which is not subject to similar export controls. “We do not agree that the discovery of a narrow jailbreak possibility should be cause for a recall of a commercial model that has been deployed to hundreds of millions of people,” Antropic wrote. The company said it would release further technical details within 24 hours.

According to Axios, the Trump administration reportedly tried to get Anthropic to suspend release of Fable 5 before the order was received, but failed. The friction gained attention earlier this year after negotiations between Anthropic and the Department of Defense broke down and the Pentagon designated the AI ​​company as a supply chain risk. Anthropic is challenging this in court (see: Pentagon’s fight for humanity draws condemnation from former Pentagon leaders).

The export control order sparked sharp reactions across the industry. Dean Ball, an AI policy expert who briefly served in the Trump administration, criticized the move on X (formerly Twitter), calling it “just a cartoon.”

Cybersecurity researcher Peter Gilnas said Anthropic created the conditions for this outcome. “If you describe your products as munitions in every press release, the government will ultimately take your word for it,” he said. “They wrote the legal predicate themselves and called it a brand.”

The workforce aspects of this directive have also received attention. Technology expert Gaurav Agarwal posed a pointed question on LinkedIn. Are foreign nationals employed by U.S. AI companies now effectively barred from using the very models they help build, and whether those same models could eventually be rolled out to their countries of origin?

Anthropic said applying the standard industry-wide would “effectively halt the introduction of new models for all frontier model providers.” The company argued that governments should have the power to prevent unsafe deployments through a legal process that is based on transparency, fairness, and technical facts, and that this action does not meet these principles. Pro, Max, Team, and Enterprise plan customers who received free access to Fable 5 until June 22 also no longer have access.