Anthropic accidentally leaks Claude Mythos, its most powerful model yet
Anthropic left nearly 3,000 unpublished assets sitting in a publicly accessible content management system. The cache included draft blog posts, internal images, PDFs, employee details, and plans for an invite-only CEO retreat in the UK. Most notably, it contained details about an unreleased AI model called Claude Mythos.
The leak was discovered by Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge, who flagged it to Fortune. Anthropic secured the data after being notified and attributed the incident to “human error in the CMS configuration.” The company stressed it was unrelated to Claude or any of its AI tools.
So what is Claude Mythos? According to the leaked draft materials, Anthropic calls it “by far the most powerful AI model we have ever developed” and describes it as a “step change” in performance. The model has finished training and is currently in early access with select customers. Anthropic confirmed its existence to Fortune after the leak, calling it the most capable system they have built to date with notably stronger performance in reasoning, coding, and cybersecurity tasks.
The leak also revealed a new model tier called Capybara. Anthropic currently offers three tiers: Opus (most capable), Sonnet (mid-range), and Haiku (lightweight). Capybara would sit above Opus as the company largest and most intelligent tier.
The cybersecurity angle is what makes this more than a standard product leak. The draft blog post warned that Claude Mythos could create “unusually high cybersecurity risk” because its offensive cyber capabilities appear to be advancing faster than defenders can keep up. Anthropic wrote that it wanted to “act with extra caution” before releasing Capybara given the potential for misuse.
This lands hard for Anthropic specifically. The company has built its brand on AI safety and responsible scaling. In February it published version 3.0 of its Responsible Scaling Policy, stating that stronger models demand stronger safeguards. It launched a transparency hub and expanded public-facing processes around trust and coordinated vulnerability disclosure. Having sensitive pre-release model details and risk assessments sitting in a publicly searchable cache undermines that narrative.
The market reacted. Cybersecurity stocks slid on Friday following the reports, with investors already on edge about Anthropic February launch of Claude Code Security, an AI tool designed to autonomously hunt software vulnerabilities. That announcement alone triggered a sell-off in traditional security vendors. The Mythos leak compounded those fears.
Anthropic said the exposed materials were early drafts that did not involve core infrastructure, AI systems, customer data, or security architecture. The company has not published a full incident timeline, the total scope of exposed files, or whether anyone beyond journalists accessed the cache before it was locked down.