Common Questions

What is Project Glasswing and who is in it?

Project Glasswing is Anthropic's restricted-access programme for Claude Mythos Preview. As of April 2026 it has 17 vetted partners, each receiving $100M in usage credits. Named partners include JPMorgan Chase, Broadcom, CrowdStrike, Nvidia, Palo Alto Networks, and Cisco.Source: ai-jobs-power-money

Why did Anthropic not release Mythos publicly and create Glasswing instead?

Anthropic withheld Mythos from public release because the model can autonomously chain 32-step cyberattacks equivalent to 20 human hours of work, posing offensive cybersecurity risk if broadly accessible. Glasswing restricts access to vetted organisations.Source: ai-jobs-power-money

Does the US government oversee Project Glasswing?

There is no publicly confirmed government oversight of Glasswing. The Treasury-Fed emergency bank meeting on 8 April 2026 suggests regulators were reacting to the programme rather than participating in its design.Source: ai-jobs-power-money

Background

Project Glasswing is Anthropic's restricted-access programme for Claude Mythos Preview, launched on 8 April 2026. It distributes the model exclusively to vetted partner organisations alongside $100 million in model usage credits each, with JPMorgan Chase publicly identified as one partner. Following Anthropic's 7 April system card, the partner list expanded by five to seventeen organisations, adding Broadcom, CrowdStrike, NVIDIA, Palo Alto Networks, and Cisco. The programme's structure (vetted partners, credit allocations rather than open pricing, and no public release) represents an unusual approach to frontier model deployment, driven by Anthropic's assessment that Mythos poses unacceptable offensive cybersecurity risk if widely accessible.

The Glasswing name draws on the glasswing butterfly, whose transparent wings make it difficult for predators to track, an apt metaphor for a programme designed to distribute powerful capability while maintaining opacity about its full partner list and usage conditions. The 7 April Alignment Risk Update marked a methodology shift: Anthropic abandoned ASL (AI Safety Level) capability benchmarks in favour of autonomy-focused threat models, meaning the criteria partners must demonstrate they can handle have changed alongside the expanded list.

The programme's existence became publicly significant when the US Treasury and Federal Reserve convened an emergency meeting with five bank CEOs specifically to discuss Mythos; JPMorgan's CEO, already a Glasswing partner, notably absent. The UK's AISI subsequently published an independent evaluation confirming a 32-step attack chain equivalent to 20 human work-hours. On 12 June 2026, the arc reached a new stage when Commerce Secretary Lutnick issued a directive forcing Anthropic to disable Claude Fable 5 and Claude Mythos 5 globally, citing a reported jailbreak from an unnamed rival. The suspension applied to both models, effectively removing Glasswing partner access pending restoration. Glasswing now sits at the centre of a live policy question: whether government should pre-authorise which organisations may receive frontier AI access and under what conditions.