Initial users of Anthropic’s Mythos continue to have access following a US directive.

      When Washington instructed Anthropic to terminate access to its most sophisticated AI model, not everyone was affected. Certain organizations that Anthropic initially selected to test Mythos retained their access to a preview of the system, despite other versions being deactivated under a US export directive, according to a Bloomberg report.

      The specifics are crucial. Anthropic restricted its Mythos Preview to about 200 organizations, including the US government, under its Glasswing program, after the model showcased a unique capability: it detected thousands of software vulnerabilities. A less advanced version of Mythos was made available more broadly but was subsequently disabled when the Commerce Department mandated Anthropic to stop access for all foreign nationals due to national-security concerns. However, the preview survived for some.

      Two confirmed firms stated they still had access. Dragos, the industrial-cybersecurity company that Accenture is in the process of acquiring a majority stake in, along with Cisco Systems, both informed Bloomberg that they maintained their Mythos Preview access.

      This detail is significant: the organizations that have kept their access are specifically those using the model for defensive security purposes, which is the primary use case Anthropic cites in advocating that frontier AI in vetted hands is a public benefit rather than a danger.

      The contrast with Europe is stark. The European cybersecurity agency ENISA, which had been invited to participate in Glasswing prior to Washington’s interference, was notified on Friday that it would no longer have access. This reversed the arrangement discussed in a meeting just days before, and it clearly answered the question that had remained from that discussion: whether a European agency could be reconciled with a US export order. The answer, for the time being, is no.

      This situation highlights the extent of discretion that Anthropic has in the void left by the directive. The export order focuses on foreign nationals, but it does not seem to specify which existing Glasswing members retain access to the preview. It was unclear how Anthropic determined individual access, implying that the company is essentially making case-by-case decisions about who can utilize one of the most advanced security models available.

      This scenario places any vendor in an uncomfortable position, and it is revealing. Throughout the year, Anthropic has cautioned against the risks posed by advanced AI while contending that models like Mythos should be accessible to defenders. The directive compelled it to make choices, and the trend of who retained access—US security firms in and a European agency out—reflects the contours of the export rule rather than an evaluation of merit.

      The model's capabilities are genuinely powerful, which is precisely the issue. A system capable of identifying thousands of vulnerabilities is equally beneficial to an attacker as to a defender, and that dual-use aspect is why Commerce sought to impose restrictions. The same functionality that prompts Dragos and Cisco to desire ongoing access is also what creates apprehension in Washington about who else could obtain it.

      The contrasting outcomes also raise questions about who controls access to dual-use AI. When a single company decides, within the framework left by a government directive, which security teams maintain access to a frontier model and which do not, practical authority over a tool pertinent to national security resides with a private entity rather than a public body.

      Granting so much discretion to one firm is considerable, and the inconsistency between the retained US firms and the eliminated European agency highlights the level of discretion in a way that warrants examination. Anthropic has not disclosed its access criteria, and the situation remains dynamic as the company navigates between governmental directives and its customers. What is evident is that the shutdown was not total; the entities that survived are predominantly US defensive-security users, and the first European participant has now been the first to be excluded.