Anthropic Previews Mythos, Its Most Powerful Model Yet. Also Leaked 500,000 Lines of Code.
Anthropic released a preview of Mythos on April 7, 2026. The model, previously codenamed Capybara, is a general-purpose system larger and more capable than the company's existing Opus lineup. Anthropic's internal assessment: "by far the most powerful AI model we've ever developed." Same day, Anthropic confirmed it had accidentally exposed nearly 2,000 source code files and over 500,000 lines of proprietary code via a bug in Claude Code 2.1.88. Launch days are complicated.
What Mythos Is
Mythos targets software coding, academic reasoning, and cybersecurity, with performance that reportedly exceeds current public models across all three. It was not specifically trained for cybersecurity. The performance comes from general capability, not narrow fine-tuning.
The model is described as general-purpose, built for Claude AI systems with strong agentic coding and reasoning. The Capybara codename is retired.
Project Glasswing
Anthropic is deploying Mythos through Project Glasswing, a security initiative with 12 core partner organizations. Named partners include Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. In total, 40 organizations will gain preview access.
The objective is defensive: scan first-party and open source software for code vulnerabilities. Results so far are notable. In recent weeks, Mythos identified thousands of zero-day vulnerabilities. Many are rated critical. Many are one to two decades old.
The age of the findings is the more interesting number. Discovering new bugs in new code is expected. Surfacing critical zero-days from the early 2000s in production systems suggests both genuine detection capability and a long period of collective neglect. The model was not specifically trained for this work. It found them anyway.
The Other Announcement
On the same day Anthropic previewed its most capable model for security work, the company confirmed the Claude Code 2.1.88 source code exposure. Nearly 2,000 files and over 500,000 lines of code were surfaced publicly. Cleanup required taking down thousands of GitHub repositories.
The timing is not ideal for a company positioning itself as a security partner to Amazon, Apple, Microsoft, and a dozen others. No indication Mythos was asked to audit the Claude Code release pipeline beforehand.
Source: Techcrunch