The Trump administration halts the rollout of Anthropic's Mythos.

      The Trump administration has expressed its opposition to Anthropic's plan for a controlled rollout of its advanced AI, which has capabilities for cyberattacks, that would extend access to about 120 organizations. This comes as the White House simultaneously considers an executive order to facilitate Anthropic's reintegration into government use. According to Bloomberg, an anonymous administration official indicated that the White House disapproves of Anthropic's intention to broaden access to its cybersecurity AI model, Mythos, by about 70 additional companies. This development was originally reported by the Wall Street Journal, and Anthropic has chosen not to comment.

      Mythos, revealed in early April as part of Anthropic's Project Glasswing initiative, is a model that can independently identify and exploit vulnerabilities in a variety of critical software, a capability Anthropic has deemed too risky for general distribution. Instead, the company has limited testing to a select number of organizations. The Trump administration has raised objections to expanding the initial group from around 50 to 120 organizations on two main points: concerns about the potential for misuse and worries about Anthropic's computing power being insufficient to support that many entities without hindering the government's effective use of the model.

      The concerns about computing power are significant. One reason for Anthropic’s current $900 billion fundraising effort is to secure adequate infrastructure to operate Mythos on a large scale. The National Security Agency is currently utilizing Mythos, and the White House seems to worry that widening access would compete with the government’s access.

      Complicating the rollout is a security breach that occurred on the same day Anthropic announced its limited release plan, where a small group of unauthorized users accessed Mythos in a private online forum. The specifics of the breach are unclear, but it highlighted the challenges of controlling access to a model of such power and heightened government fears about expanding its user base.

      The capabilities of Mythos are well-documented, having autonomously identified thousands of zero-day vulnerabilities across major operating systems and web browsers during testing, achieving 73% success in expert-level cybersecurity challenges, and becoming the first model to complete a 32-step simulated corporate network attack from start to finish. These aren’t just theoretical dangers; they are established capabilities that the US government has recognized as a security threat and has seemingly aimed to monopolize through NSA access.

      The White House's objection is notable, especially as it coincides with ongoing efforts to draft an executive action that would allow government agencies to bypass the Pentagon’s classification of Anthropic as a national security supply chain risk, which was established in early 2026 after unsuccessful negotiations regarding the military's potential use of Claude for autonomous weapons and domestic surveillance—uses that Anthropic’s CEO Dario Amodei has publicly opposed.

      This week, the White House is bringing together companies from various sectors to inform this draft executive action and is also conducting discussions with Anthropic while signaling its disapproval of the Mythos expansion plan. Both White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent held what was described as a productive introductory meeting with Amodei earlier this month.

      The official stance from the White House claims to balance innovation with security while collaborating with the private sector; however, this does not clarify the apparent contradictions in their approach. The dispute over Mythos illustrates a complex regulatory landscape for a company whose AI model is being utilized by the NSA, contested by the Pentagon, pursued by the White House for re-entry into federal use, and now restricted from extending access to civilian companies initially approved. The outcomes of this week’s discussions will shape not only Anthropic’s rollout strategies but also the broader framework regarding how the US government plans to regulate AI models capable of offensive cybersecurity actions.