AIAcademy · AIAcademy · 2026-05-16
Read about Project Glasswing on anthropic.com
For the first time, a frontier lab has shipped a model class it is not selling. Anthropic's Mythos Preview — a tier above Opus, internally codenamed Capybara — reaches the world only through Project Glasswing, a critical-infrastructure pilot with roughly 50 organizations: AWS, Apple, Cisco, JPMorgan, Microsoft, Nvidia, CrowdStrike, the Linux Foundation, and ~40 others. No public API. No consumer product. ASL-3 protections applied throughout.
The gating concern is concrete. Mythos found a 17-year-old FreeBSD NFS remote-code-execution bug fully autonomously and clears 93.9% on SWE-bench Verified. Anthropic's own framing — "currently far ahead of any other AI model in cyber capabilities" — describes both the marketing pitch and the reason the model is not on a price list. Bruce Schneier's reading of the launch is that we have entered the era of dual-use defense models: capabilities good enough to harden critical infrastructure are, by construction, capabilities good enough to attack it.
The pattern matters beyond Anthropic. The Fortune scoop that surfaced Mythos in March made clear the company had already chosen not to ship the model normally — staged release was the plan, not the fallback. The Opus 4.7 launch note explicitly frames future consumer Opus as Mythos-with-safeguards. Read together: for the next 24 months, expect a growing class of frontier models that exist, work, and are deployed — but never become commodity APIs.
The contested claim worth carrying around: is this a responsible deployment model, or a regulatory-capture vehicle that locks frontier capability behind a curated client list? Both readings are live.