Project Glasswing: Anthropic Turned Its Scariest Model Into a Service
Anthropic got Amazon, Apple, Google, Microsoft, and eight other tech giants into the same room this week. Not to talk about standards. To deploy an AI agent that finds zero-day vulnerabilities faster than any human team.
Project Glasswing launched April 7 around Claude Mythos Preview, a model Anthropic considers too capable to release publicly. It has already found thousands of high-severity zero-days across every major operating system and web browser -- bugs that survived decades of human code review and millions of automated tests. The benchmark numbers back it up: 83.1% on vulnerability reproduction versus 66.6% for Claude Opus 4.6.
The real story is not the model. It is the business model. Anthropic is committing $100 million in usage credits, plus $4 million to open-source security foundations. The 12 founding partners -- AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks -- plus about 40 critical infrastructure organizations get exclusive access. Nobody else.
CrowdStrike and Palo Alto stocks dropped 6-7% when Mythos capabilities leaked last week. Now they are inside the tent rather than outside it. That is the calculation everyone made: if this model is going to reshape cybersecurity anyway, better to be at the table.
Anthropic will report findings publicly within 90 days. This is the first time a model company has said our AI is too dangerous to release, so we will operate it as a managed service for critical infrastructure. It is either responsible stewardship or the cleverest moat in AI. Probably both.
https://www.anthropic.com/glasswing
← Back to all articles
Project Glasswing launched April 7 around Claude Mythos Preview, a model Anthropic considers too capable to release publicly. It has already found thousands of high-severity zero-days across every major operating system and web browser -- bugs that survived decades of human code review and millions of automated tests. The benchmark numbers back it up: 83.1% on vulnerability reproduction versus 66.6% for Claude Opus 4.6.
The real story is not the model. It is the business model. Anthropic is committing $100 million in usage credits, plus $4 million to open-source security foundations. The 12 founding partners -- AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks -- plus about 40 critical infrastructure organizations get exclusive access. Nobody else.
CrowdStrike and Palo Alto stocks dropped 6-7% when Mythos capabilities leaked last week. Now they are inside the tent rather than outside it. That is the calculation everyone made: if this model is going to reshape cybersecurity anyway, better to be at the table.
Anthropic will report findings publicly within 90 days. This is the first time a model company has said our AI is too dangerous to release, so we will operate it as a managed service for critical infrastructure. It is either responsible stewardship or the cleverest moat in AI. Probably both.
https://www.anthropic.com/glasswing
Comments