Anthropic's Claude Mythos is So Good the Company Won't Let You Use It

Anthropic has begun testing a new advanced AI model, Claude Mythos Preview, but is restricting access to a limited group of partners due to cybersecurity concerns.

The model, described as a general-purpose frontier system, was first revealed through a misconfiguration in the company's internal content system in late March. The unpublished material indicated that Claude Mythos would be a new tier of models more capable than its current flagship, Opus.

Rather than releasing the model publicly, Anthropic is providing access through an initiative called Project Glasswing. Under this program, a small group of organizations will use the model for defensive security work.

Launch partners include Amazon, Apple, Broadcom, Cisco, CrowdStrike, Linux Foundation, Microsoft, and Palo Alto Networks, along with around 40 additional organizations.

These participants will use the model to scan and secure their systems as well as open-source tools.

Although not specifically designed for cybersecurity, Claude Mythos shows strong performance in coding and reasoning tasks. On the CyberGym benchmark for vulnerability analysis, the model scored 83.1%, compared with 66.6% for Opus 4.6, which previously led the rankings.

Anthropic said the model has already identified thousands of zero-day vulnerabilities, including critical issues. One example cited was a previously unknown bug in OpenBSD that had remained unpatched for 27 years. The model also demonstrated the ability to combine multiple vulnerabilities in the Linux kernel to gain superuser access.

Anthropic said it is taking a cautious approach to releasing the model, citing potential near-term cybersecurity risks. The company aims to give defenders time to strengthen systems before similar capabilities become widely available.

CrowdStrike CTO Elia Zaitsev said the time between discovering and exploiting vulnerabilities has significantly shortened with AI, noting that such tools can operate at scale.

The initiative also targets open-source ecosystems, where security resources are often limited. According to Jim Zemlin, the use of AI could help maintainers identify and fix vulnerabilities more efficiently.

Anthropic is supporting the effort with $100 million in usage credits for participating organizations and $4 million in donations to open-source security groups.

Anthropic said it plans to release Mythos-class models more broadly in the future. For now, access remains limited to Project Glasswing participants as the company evaluates risks and potential impacts.