Anthropic on Tuesday launched Project Glasswing, a cybersecurity coalition built around an unreleased frontier model called Claude Mythos Preview that the company says has already found thousands of zero-day vulnerabilities across every major operating system and web browser. AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, JPMorganChase, Cisco, Broadcom, Palo Alto Networks, and the Linux Foundation are the launch partners, with over 40 additional organizations getting access.
Anthropic is committing up to $100 million in usage credits and $4 million in direct donations to open-source security organizations. The model will not be made publicly available.
The backstory nobody planned
Mythos wasn't supposed to be announced this way. A data leak on March 27 exposed a draft blog post describing the model, internally codenamed "Capybara," as a new tier above Opus. Anthropic confirmed its existence but attributed the exposure to a CMS misconfiguration. Days later, the company accidentally leaked Claude Code's source code through a botched npm release, then knocked down roughly 8,100 GitHub repos with an overly broad DMCA takedown while trying to clean up.
Two major security lapses in five days from a company that sells safety as a core brand value. And now that same company is asking the industry to trust it with coordinated vulnerability disclosure at scale.
What Mythos actually found
The headline numbers are striking, if you take them at face value. According to Anthropic's red team blog, Mythos found a 27-year-old vulnerability in OpenBSD that allowed remote crashes, a 16-year-old bug in FFmpeg that automated testing tools had hit five million times without catching, and chained several Linux kernel vulnerabilities to escalate from user access to full control. All three have been patched.
On CyberGym, a vulnerability reproduction benchmark, Mythos scored 83.1% versus Claude Opus 4.6's 66.6%. On SWE-bench Verified, the gap is wider: 93.9% to 80.8%. These are large jumps, though CyberGym is relatively new and the evaluation methodology deserves scrutiny that hasn't happened yet in public.
"The window between a vulnerability being discovered and being exploited by an adversary has collapsed," CrowdStrike CTO Elia Zaitsev said in a blog post, which is the kind of statement CrowdStrike CTOs are paid to make. More interesting is what Microsoft's Igor Tsyganskiy added: when tested against CTI-REALM, Microsoft's open-source security benchmark, Mythos showed "substantial improvements" over previous models. An independent confirmation carries more weight than Anthropic grading its own homework.
The disclosure problem
Over 99% of the vulnerabilities Mythos found haven't been patched yet. Anthropic says it has published cryptographic hashes of the details and will reveal specifics after fixes are in place. It has also hired professional security contractors to manually validate every report before sending it to maintainers, claiming 89% exact agreement with Claude's severity assessments across 198 reviewed reports.
But here's the tension VentureBeat raised: flooding open-source maintainers, many of them unpaid volunteers, with an avalanche of critical bug reports could do more harm than good. The Linux Foundation's Jim Zemlin framed Glasswing as leveling the playing field for under-resourced maintainers, but dumping thousands of vulnerability reports on people who maintain code in their spare time is not, on its own, a solution. Anthropic's $2.5 million donation to Alpha-Omega and OpenSSF through the Linux Foundation, plus $1.5 million to the Apache Software Foundation, is a start. Whether it's enough depends on the actual volume of work headed their way.
So what does it cost?
After the $100 million in credits runs out, Mythos Preview will be available to participants at $25/$125 per million input/output tokens. That's roughly five times the cost of Opus. Partners can access it through the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry.
Anthropic says it plans to launch new safeguards with an upcoming Claude Opus model, using it as a testing ground before eventually bringing Mythos-class capabilities to a wider audience. Security professionals affected by those safeguards will be able to apply to a forthcoming Cyber Verification Program. The company has committed to reporting publicly on lessons learned within 90 days.
The timing is pointed. Anthropic has been in discussions with CISA and the Center for AI Standards and Innovation about Mythos's capabilities. The Glasswing announcement explicitly frames this as a national security issue, arguing that maintaining a lead in AI is essential for democratic states. That's a lobbying position dressed up as a product launch, though the underlying concern about proliferation is real enough.
