Anthropic Restores Mythos 5 Access for US Organizations

Commerce Secretary Howard Lutnick sent a letter to Anthropic on June 26, authorizing a limited restoration of access to the Claude Mythos 5 model, the company's top cybersecurity-focused AI. The letter, seen by Axios, NBC News, and CNBC, states: 'a license will no longer be required to export, reexport, or in-country transfer the Claude Mythos 5 Model to entities identified in Annex A to this letter and their foreign national employees, or to Anthropic's foreign national employees.' Semafor first reported on the letter. Multiple outlets report that Annex A covers roughly 100 vetted organizations including government agencies and private firms (NBC News; Reuters; CNBC; Semafor). Fable 5 remains under export-control restrictions for all users (Axios; Reuters).

The June 12 directive, reported by Axios and Reuters, was prompted by concerns - raised internally by Amazon and communicated to Treasury Secretary Scott Bessent - that Anthropic's frontier models could be jailbroken for malicious purposes. Anthropic dispatched a team of scientists and engineers to work with Commerce Department and Office of the National Cyber Director counterparts. The Lutnick letter cites 'significant progress' in those talks and states that Anthropic committed to work with the government on 'protocols and standards and releases for the Covered Models,' without elaborating on what that entails (Axios; Semafor).

Lutnick explicitly reserved the right to revoke or adjust the access list: 'I reserve the right to reevaluate and adjust the scope of license requirements on the Covered Models, should circumstances change,' the letter states (Axios). The Annex A list of approved entities has not been published, and the specific technical mitigations Anthropic provided to satisfy the government's diversion-risk assessment have not been disclosed (Axios; Semafor). Reporting also notes that OpenAI was simultaneously subjected to government reviews of its own frontier models, suggesting coordinated oversight across frontier labs rather than an Anthropic-specific event (Reuters; Axios).

This episode establishes a precedent pattern: export-control tools applied to frontier AI models create a formal, non-transparent vetting process where approved entities gain preferential access. For practitioners, this means operational friction on wide public releases of high-capability models - especially those with dual-use security capabilities - and growing compliance requirements for teams deploying such models in regulated or government-adjacent environments.

An August cybersecurity executive order deadline calls for federal agencies to create a formal process for assessing AI models' cyber capabilities. Anthropic is also in discussions with government officials about a formalized policy framework to address national-security concerns before models are released, per a source familiar with the talks (Axios). Whether that framework, once published, removes the ad hoc export-control approach or codifies it will shape how all frontier labs plan future model releases.