US Government Allows Limited Use of Anthropic's Claude Mythos 5 Cybersecurity Model

The US government has approved Anthropic to resume limited access to its advanced cybersecurity AI model, Claude Mythos 5, after previously ordering a block just two weeks ago. This new authorization permits over 100 American companies, government agencies, and organizations to use the model, which is regarded as one of Anthropic's most powerful tools for cybersecurity tasks. The initial ban was imposed due to security concerns and allegations that the model bypassed protective mechanisms in advanced AI systems.

The easing of restrictions does not extend to the general public, and another model, Claude Fable 5, remains inaccessible. US Secretary of Commerce Howard Lutnick sent an official letter to Anthropic leadership confirming that appropriate safeguards have been established to allow trusted partners to utilize Claude Mythos 5. The approval primarily targets entities involved in protecting critical infrastructure, including government agencies and American firms in security, energy, communications, and cybersecurity sectors.

As part of the agreement, the government exempted the model from a recent export control order, enabling non-US citizen employees at Anthropic and approved organizations to access the model under new oversight rules. This policy shift followed pressure from US security agencies such as the NSA, which had lost access to Anthropic's cybersecurity tools. Technology industry representatives warned that overly strict restrictions could undermine US companies' competitiveness against China in artificial intelligence.

Anthropic confirmed efforts to expedite access for approved organizations and pledged ongoing collaboration with the government to develop protocols, standards, and release procedures for advanced AI models. No timeline has been set for opening Claude Mythos 5 to the broader public or additional companies, with future decisions dependent on further review by the US Department of Commerce and government risk assessments.