Anthropic announced Claude Mythos Preview this week, a new frontier model that’s more capable than Claude Opus 4.6 - and which you almost certainly won’t be able to use anytime soon.
The reason isn’t cost or readiness. It’s that the model is genuinely scary good at finding and exploiting security vulnerabilities.
Mythos scored 93.9% on SWE-bench Verified (Opus 4.6 scored 80.8%), 77.8% on SWE-bench Pro (Opus 4.6 scored 53.4%), and 83.1% on the CyberGym benchmark versus Opus 4.6’s 66.6%. Those are big jumps. But the benchmark numbers aren’t really the story.
The story is what the model actually did when pointed at real software. It found tens of thousands of zero-day vulnerabilities across major operating systems and browsers - many of them sitting undetected for one to two decades. A 27-year-old bug in OpenBSD. A 16-year-old bug in FFmpeg. Vulnerabilities in the Linux kernel. For comparison, Opus 4.6 found around 500. Mythos found tens of thousands.
And it doesn’t just find the bugs. It writes working exploits too. That’s apparently what pushed Anthropic to handle this differently.
Anthropics head of frontier red team said this is “the first time a model is this good that we decided to approach release in a very different way.” The 244-page system card is worth reading if you have time.
Instead of a public launch, Anthropic created Project Glasswing - a restricted cybersecurity initiative with 40+ hand-picked partner companies using Mythos for defense only. The list includes Amazon, Apple, Microsoft, Google, Nvidia, Broadcom, Cisco, CrowdStrike, Palo Alto Networks, JPMorgan, and the Linux Foundation. Anthropic is providing up to $100 million in usage credits to these partners and $4 million to open-source security organizations. They’re also briefing CISA and the Commerce Department.
The stated goal is eventually enabling broader deployment of Mythos-class models safely. Whether that happens, and when, remains unclear.
This is the first time an AI company has held back a model specifically because of capability concerns rather than technical or commercial ones. Depending on your perspective, that’s either the most responsible thing an AI company has done yet, or a sign that things are moving faster than anyone is comfortable admitting.
There’s also a real tension in the access question. The companies in Project Glasswing - already among the most powerful in tech - now have exclusive use of what Anthropic is calling its most capable model. That’s a meaningful advantage, and it’s not one that’s being distributed evenly. Some people will find that reassuring. Others will find it troubling. Both reactions seem reasonable.

Collections

Anthropic announced Claude Mythos Preview, a frontier model significantly more capable than Claude Opus 4.6, but chose not to release it publicly due to its unprecedented ability to find and exploit security vulnerabilities. The model scored 93.9% on SWE-bench Verified and discovered tens of thousands of zero-day vulnerabilities across major operating systems and browsers — including bugs decades old — compared to ~500 found by Opus 4.6. It also writes working exploits. Instead of a public launch, Anthropic created Project Glasswing, a restricted initiative giving access to 40+ hand-picked companies including Amazon, Apple, Microsoft, Google, and Cisco for defensive use only. Anthropic is providing up to $100M in usage credits to partners and $4M to open-source security organizations, while briefing CISA and the Commerce Department. This marks the first time an AI company has withheld a model specifically due to capability concerns, raising questions about equitable access and the pace of AI advancement.

Anthropic launches Claude Mythos Preview in restricted partner program