Anthropic's Unreleased "Claude Mythos" AI Model Leaked Due to Configuration Error

Anthropic's unreleased artificial intelligence model, "Claude Mythos," was inadvertently exposed online following a configuration error in the company's content management system. The leak revealed details of the model, codenamed "Capybara," which is described as a new top-tier offering surpassing the capabilities of the current Claude Opus 4.6 in programming and reasoning.

Server rack with an open bay and digital error message, symbolizing a configuration error.

The exposure occurred when approximately 3,000 internal documents, intended to be confidential, became publicly accessible through a database error linked to an external tool used for publishing official blog posts. Cybersecurity firm LayerX Security and researchers from Cambridge University identified the vulnerability.

Fortune retrieved a draft blog post detailing Claude Mythos, which reportedly completed training but has not been released due to internal evaluations deeming it too dangerous. The draft indicates Mythos possesses powerful "cyberattack" and "defense capabilities," with a potential parameter scale of 10 trillion. Internal tests suggest it could introduce unprecedented security risks.

Digital brain over a cracked circuit board, symbolizing powerful AI with security risks.

Anthropic's Claude models typically come in three sizes: Opus, Sonnet, and Haiku. Capybara is positioned as a larger, smarter, and more expensive "top-tier" model than Opus. The leaked draft highlighted Mythos's superior performance in software coding, academic reasoning, and cybersecurity compared to Claude Opus 4.6.

Cybersecurity Risks and Controlled Release

The draft blog post warned that Claude Mythos presents significant cybersecurity risks. It states that Mythos is "far ahead of any model in the world" in terms of cyberattacks, raising concerns that it could be exploited by hackers to launch large-scale, highly destructive attacks that exceed current defensive capabilities.

Digital firewall under attack, symbolizing cybersecurity risks and defense.

Due to these risks, Anthropic's release plan for Mythos will initially focus on cyber defenders. The company intends to offer early access to relevant organizations, enabling them to enhance the robustness of their codebases against future AI-driven exploits.

Executive Meeting Details Also Exposed

Beyond the new model, the leak also included a PDF file detailing a confidential business itinerary for Anthropic. This included plans for CEO Dario Amodei to host an invite-only summit for top European business leaders at a luxury hotel in the UK. The event was intended to showcase "Claude's unreleased mysterious capabilities" to potential major clients.

Elegant conference room in a luxury hotel, set for a high-level business summit.

Anthropic has acknowledged the leak, attributing it to a "human configuration" error within an external content management system tool. The company stated that many of the exposed files were leftover or discarded materials from past blog posts, such as images and banners, but also included clearly private or internal documents, including one related to an employee's parental leave. Anthropic has since blocked the data retrieval channel.