Anthropic has been cleared by the US government to restore limited access to its powerful Mythos 5 AI model, after national security concerns forced the company to shut off access to two of its most advanced systems earlier this month.

The approval allows Mythos 5 to be redeployed to a small group of trusted partners, including cyber defenders and infrastructure providers, but the related Fable 5 model remains restricted.

The decision follows two weeks of talks between Anthropic and the US Commerce Department after officials raised concerns that safety guardrails on Mythos 5 and Fable 5 could be bypassed, potentially allowing the models to be misused.

Commerce Secretary Howard Lutnick (pictured) reportedly told Anthropic that the company had made “significant progress” in addressing risks linked to the models, allowing access to resume for selected users.

Anthropic said Mythos 5, which it describes as its strongest cybersecurity model, would be restored for approved organisations as quickly as possible.

US Senate confirms Howard Lutnick as commerce secretary

“We are pleased to see this progress and continue to work with the government to expand access to Mythos 5 and make Fable 5 available for general use again,” the company said.

Mythos 5 is designed for cybersecurity work, including finding software vulnerabilities. Before the government intervention, it had reportedly been made available through an invite-only program to around 200 firms, including major technology companies such as Apple, Google, Cisco, Nvidia and Microsoft.

The move marks one of the most significant examples so far of the US government intervening directly in the rollout of a frontier AI model.

It also highlights growing tension between AI developers, regulators and national security agencies as more powerful models move beyond chatbots and into areas such as cybersecurity, infrastructure protection and defence-related work.

The Anthropic decision has wider implications for the AI industry, with OpenAI also reportedly limiting the release of its latest GPT-5.6 model to selected government-approved partners.