Samuel Boivin/NurPhoto via Getty Images

• Industry
• Commerce

The Trump administration told Anthropic late Tuesday it is lifting export controls on the Mythos 5 and Fable 5 models allowing the company to restore broader access to some of its most powerful artificial intelligence systems and ending a closely watched standoff in the world of U.S. AI governance.

The shift lets Anthropic reopen access for users in the U.S. and overseas, after the company notably agreed to steps meant to limit chances for the Fable 5’s cyber capabilities to be abused. Both the company and the Commerce Department confirmed the decision Tuesday night.

“We’ll begin restoring access tomorrow, and will share an update soon,” Anthropic said in a post on X. “We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.”

The export-control order was invoked on national security grounds but confused much of the AI and cybersecurity community, which raised questions about what specific characteristics made the models uniquely risky.

Earlier this year, Anthropic released its first Mythos variant through Project Glasswing, an ongoing limited-access initiative designed to put its powerful cyber-AI models in the hands of trusted organizations for mainly cyberdefense purposes. The company has since worked to expand Mythos access across the U.S. government, though the rollout has at times been rocky, including for a key cybersecurity agency deemed a clear candidate to use the systems.

Fable 5, released in early June, is the broader-access, safeguarded version of Anthropic’s Mythos-class model, while Mythos 5 is the restricted Project Glasswing version made available to vetted groups with some cybersecurity safeguards removed. 

The NSA, which was using the latest Mythos build, was among those affected by the export control order. This past Friday, the Trump administration partially lifted that ban on Mythos 5, allowing a select group of around 100 organizations to regain access, but kept in place restrictions on Fable 5.

The reversal is a major reprieve for Anthropic and many of the cyberdefenders it has been trying to equip. It also comes as Chinese open-source models are beginning to show similar cyber capabilities to that of major U.S. AI labs.

“Stronger AI models aren’t a genie that can be crammed back into a bottle, no matter how much we might like to,” ThreatLocker CEO Danny Jenkins said in a statement. “All of this means that the only ones being restricted by the export controls are organizations that desperately need to test their systems and code.”

The government’s concern over Fable 5 was tied to an Amazon report — the significance of which Anthropic disputed — that the public became aware of around the time the controls were imposed. In a blog post issued Tuesday night, Anthropic acknowledged that the report’s researchers claimed to have found a way to bypass Fable 5’s safeguards and get it to identify software flaws that could be exploited.

But the company contends the issue did not show Fable 5 was giving general users access to the more sensitive cyber capabilities reserved for Mythos. Still, Anthropic said in the blog that it trained a new safety filter which blocks the reported workaround more than 99% of the time.

Anthropic also said it is working with Amazon, Microsoft, Google and other Glasswing partners on a common framework for assessing AI jailbreaks, including when a bypass is serious enough to require new safeguards or other action from model developers.

The AI company is already in an ongoing legal fight with the Pentagon over its designation as a “supply-chain risk,” which the company has argued was retaliation for its refusal to relax limits on certain military uses of its models. A federal judge granted Anthropic a preliminary injunction in March blocking parts of that designation and a subsequent order to end all government use of the company’s products, but the government appealed, and litigation has continued into June.

The lack of clarity around the export directive has also shaped how other AI firms are approaching similar model deployments. OpenAI said Friday it will initially limit access to three of its GPT-5.6 models after conversations with government officials, while it tests the systems with select partners before a broader release. 

The company notably said it does not want government access reviews to become the long-term default, though called the short-term preview the best path toward wider availability as it works with the administration on a future release framework. A sweeping June executive order called for a voluntary framework giving the government early access to some frontier models for up to 30 days before they are released to other trusted partners.