Detailed Analysis
Approximately 200 organizations enrolled in Anthropic's Project Glasswing program have retained access to Mythos Preview following a US government order that suspended broader public and commercial availability of Fable 5 and Mythos 5, according to Bloomberg reporting from June 19, 2026. The carve-out represents a significant distinction between general market access and specialized, vetted partner access, with high-profile members including Cisco, Amazon Web Services, and JPMorgan Chase & Co. among those permitted to continue their use of the advanced AI systems. The selective preservation of access underscores how Anthropic has structured its most sensitive AI deployment through a tiered framework that insulates critical research partnerships from sweeping regulatory actions.
Project Glasswing appears to function as Anthropic's curated early-access program for cybersecurity-focused partners, specifically designed to facilitate vulnerability research and adversarial testing of advanced AI systems. The program's explicit orientation toward security testing suggests its participants are engaged in work that regulators and Anthropic itself likely consider beneficial or necessary, even in a restricted-access environment. By maintaining Mythos Preview access for these roughly 200 entities while halting Mythos 5 and Fable 5 more broadly, Anthropic effectively preserved a controlled research channel that may serve both defensive AI safety goals and government interests in understanding the capabilities and risks of frontier models.
The US government's shutdown order affecting Fable 5 and Mythos 5 reflects a pattern of increasing regulatory intervention in frontier AI deployment that has accelerated as models grow more capable. The fact that a federal order could halt broad access to specific named model versions indicates a maturing regulatory apparatus with the authority and specificity to target particular AI products, a notable evolution from earlier, more diffuse policy discussions. The carve-out for Project Glasswing members suggests regulators and Anthropic negotiated or pre-established exceptions for research-oriented use cases, which implies some degree of coordination between the company and government authorities prior to or concurrent with the order.
The composition of Project Glasswing's membership — spanning cloud infrastructure providers like AWS, financial institutions like JPMorgan Chase, and cybersecurity firms like Cisco — illustrates how Anthropic has cultivated a cross-sector coalition of partners that embed advanced AI testing into critical economic and security infrastructure. These are not peripheral players; they represent nodes in the financial system, internet infrastructure, and enterprise security that carry systemic importance. Their continued access to Mythos Preview, even under a broader shutdown, signals that the operational and research dependencies these organizations have developed around Anthropic's frontier models are now treated as carrying enough institutional weight to warrant protected status.
The episode reflects a broader tension in the AI industry between the pace of capability development, the breadth of commercial deployment, and the capacity of regulatory frameworks to respond in targeted rather than blunt ways. Anthropic's ability to maintain a functioning research ecosystem for 200 organizations even amid a government-mandated access halt demonstrates the strategic value of building structured partner programs with clearly defined purposes and vetted membership. As governments worldwide grapple with how to regulate increasingly powerful AI systems, the Project Glasswing model — insulating security-critical research partnerships from sweeping access restrictions — may emerge as a template that other frontier AI developers adopt to preserve operational continuity with key institutional partners during periods of regulatory uncertainty.
Read original article →