Anthropic Built an AI So Capable It Decided Not to Release It

On April 7, Anthropic announced Claude Mythos Preview, its most capable AI model to date, and said it would not be making it generally available. Instead, controlled access goes to 11 founding partners from major technology companies, under a program called Project Glasswing. The stated purpose is to let defenders find and patch critical software vulnerabilities before a model with comparable capabilities reaches the open market through a competitor.

The reason for that caution is concrete: in testing, Mythos Preview demonstrated an ability to autonomously find and exploit software vulnerabilities at a scale no prior AI system had managed. It discovered a 17-year-old remote code execution bug in FreeBSD's NFS implementation (CVE-2026-4747), a 27-year-old denial-of-service flaw in OpenBSD's TCP handling, and a 16-year-old buffer overflow in FFmpeg's H.264 codec that had survived five million hits from automated fuzzing tools without being flagged. It also generated working exploits for every major web browser and identified multiple privilege escalation chains in the Linux kernel.

The numbers are striking. Mythos achieved a vulnerability exploitation success rate of roughly 72.4%, compared to under 1% for the previous generation of frontier models. It generated 181 working JavaScript engine exploits against Mozilla Firefox; the prior top model had produced two. Anthropic estimates more than 1,000 critical-severity vulnerabilities remain unpatched in a responsible disclosure queue, with thousands more under validation. Partners have a 135-day window to disclose findings.

What Project Glasswing is

The founding partners in Project Glasswing are Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. About 40 additional organizations maintaining critical open-source software also received access. Anthropic committed $100 million in Mythos usage credits to the group and $4 million in direct funding to open-source security foundations including Alpha-Omega, OpenSSF, and the Apache Software Foundation.

Partners are required to report findings within 90 days and follow responsible disclosure standards. After their credit allocation runs out, access is priced at $25 per million input tokens and $125 per million output tokens, considerably above standard frontier model pricing.

Anthropic's official rationale: "By releasing this model initially to a limited group of critical industry partners and open source developers with Project Glasswing, we aim to enable defenders to begin securing the most important systems before models with similar capabilities become broadly available."

The asymmetry problem

Security researcher Bruce Schneier acknowledged the capability gains are real and significant. Mythos, he wrote, can chain multiple bugs, operate without human guidance, and generate working exploits far beyond what earlier models could manage. But he flagged a crucial point: finding and fixing vulnerabilities is currently easier than finding and exploiting them. That asymmetry favors defenders. The question is how long it holds as models improve and access becomes harder to restrict.

Analyst Zvi Mowshowitz noted that fewer than 1% of the vulnerabilities Mythos discovered have yet been disclosed, and that organizations with limited security budgets, hospitals, municipal governments, smaller open-source projects, face asymmetric risk if capable offensive models become broadly available before they can patch their systems.

Anthropic's own assessment was direct: "Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors committed to deploying them safely."

Whether Project Glasswing solves that problem or merely buys a few months is the question this announcement leaves open. The 135-day disclosure window means some of the discovered vulnerabilities will be patched before the model's capabilities become common knowledge. Whether that window is wide enough is harder to say.

Sources: Anthropic Red Team, Foreign Policy, TechCrunch, Fortune, Schneier on Security, NBC News