Claude Mythos: Anthropic Built a Model That Finds Zero-Days—Then Locked It Away
Anthropic released Claude Mythos.
Then locked it away.
Not unreleased. Released—but not for you.
What the Official Announcement Actually Said
According to Anthropic’s official announcement, Claude Mythos discovered thousands of zero-day vulnerabilities across every major operating system and browser in controlled testing—including the Linux kernel.
Specific examples mentioned:
- Vulnerabilities found in every major browser
- Vulnerabilities found in the Linux kernel, chained together to form complete attack paths
This capability was demand-driven. Anthropic noted that someone had already used Claude to identify vulnerabilities, develop malware, extract sensitive data, and generate tailored ransom demands.
So Anthropic made a choice: model released, but not public. Through Project Glasswing, access granted only to vetted partners.
Why This Matters
Automated vulnerability discovery is a turning point for AI security.
Before, vulnerability research depended on human security researchers—experience, time, luck. Now a model can systematically mine zero-days at a scale humans cannot match.
This changes the cost dynamics of offense vs. defense. Defenders can use AI to find vulnerabilities proactively. So can attackers.
Anthropic chose: only defenders get access.
The Problems Official Didn’t Answer
The logic of giving vulnerability discovery to defenders while restricting it from attackers sounds reasonable. But several questions Anthropic did not address:
1. Who verifies “defender” credentials
CrowdStrike and Palo Alto Networks are security vendors—and they sell security products. Letting the same people know vulnerabilities first, then sell fixes, creates a conflict of interest.
2. Why these companies specifically
The partner list includes no independent security research institutions, no academic security teams—only commercial companies. Vulnerability research成果 flow to commercial entities; independent researchers are excluded.
3. Who sets patching priority
If a vulnerability affects billions of devices but fixing it doesn’t serve any Glasswing partner’s commercial interests—will it be prioritized?
Anthropic gave no answers.
About Claude Mythos Itself
Some model parameters (from Anthropic official):
- Accessed through Project Glasswing, no public release planned
- Partners: AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia, Linux Foundation, Palo Alto Networks
Specific benchmark data and pricing—Anthropic has not published full details publicly.