Anthropic recruits Big Tech to patch the internet with unreleased Mythos model

If an AI model could effortlessly uncover a critical flaw that humans missed for 27 years, would you release it to the wild or lock it in a high-security vault?

Anthropic answered that question on April 7 by unveiling Project Glasswing, an alliance formed to defend critical software using Claude Mythos Preview—a powerful, unreleased frontier model. Instead of a general rollout, Anthropic is handing the keys to a tightly vetted circle of defenders.

The initial roster includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, alongside more than 40 other infrastructure organizations. To ensure cost is not a barrier, Anthropic committed $100 million in Mythos usage credits and $4 million in cash donations to open-source security groups. Post-preview pricing is set at $25 per million input tokens and $125 per million output tokens.

Unearthing flaws that survived millions of tests

Mythos, previously leaked under the internal codename "Capybara," sits in a tier above the Opus family. It is a general-purpose model, but its advanced agentic coding and reasoning make it terrifyingly good at cybersecurity.

Over just a few weeks, Mythos Preview autonomously discovered thousands of critical zero-day vulnerabilities across every major operating system and web browser. It worked entirely without human guidance, exposing bugs that had survived decades of scrutiny.

In OpenBSD, renowned for its hardened architecture, Mythos found a 27-year-old flaw that let remote attackers crash systems simply by connecting. In FFmpeg, a backbone of web video, it spotted a 16-year-old bug in a single line of code that automated testing tools had pinged five million times without incident. On Linux, the model autonomously chained multiple kernel bugs together to escalate ordinary user privileges into total root takeover.

The irony is that Anthropic announced this while nursing its own security black eyes. Last month, Fortune revealed the model leaked because an internal draft sat on an unsecured public data lake. Weeks later, Anthropic accidentally exposed nearly 2,000 source code files and 500,000 lines of code during a Claude Code update, inadvertently breaking thousands of GitHub repos during cleanup. Meanwhile, the lab remains locked in a legal battle with the Trump administration after the Pentagon labeled it a supply-chain risk over its refusal to support autonomous targeting or domestic surveillance.

Why it matters

Anthropic's restricted playbook makes sense, but it highlights just how uneven the AI landscape is becoming. Cybersecurity has historically favored attackers, who only need to find one unpatched crack. By turning Mythos loose on foundational systems, Anthropic is attempting to tip the scales back toward defenders before malicious actors train models with identical capabilities.

For open-source projects, this is a massive lifeline. Critical libraries like FFmpeg run the modern web on shoestring budgets. Giving maintainers access to top-tier agentic scanning, backed by $100 million in credits and $4 million in grants, fixes problems that human audits simply never had the bandwidth to catch.

Still, keeping the model behind walled gardens carries obvious friction. We think the "defenders only" approach is prudent right now, but it creates a fragile moat. If an unaligned competitor or adversary matches Mythos's capabilities, anyone outside this elite partner circle becomes an easy mark. Given Anthropic's own recent stumbles with code leaks, keeping frontier capabilities safely under lock and key is going to be far easier said than done.