Claude Mythos by Anthropic: A Leak Reveals Its Capabilities

An Accidental Revelation Shakes the AI World

The unexpected discovery of Claude Mythos, an artificial intelligence model developed by Anthropic, was caused by a simple configuration error. This incident has highlighted the considerable offensive potential of this model in cybersecurity, raising concerns about its use.

The Claude Mythos model stands out for its power and cost, surpassing its predecessors. Currently, it is in a testing phase with a limited number of clients, demonstrating its ability to exploit software vulnerabilities on an impressive scale.

Anthropic, like other major companies in the AI sector, is navigating between highlighting the potential risks of its technologies and promoting responsible use, primarily targeting cybersecurity professionals.

A Human Error Behind the Leak

On March 26, 2026, an error in the configuration of Anthropic's blog allowed the unveiling of Claude Mythos, the company's next model. The leaked document describes a model whose offensive capabilities in cybersecurity are, according to Anthropic, unprecedented.

This leak was not the result of a cyberattack or an internal disclosure, but rather a simple human error. The content management system used by Anthropic inadvertently made certain files public, including drafts of blog posts. Files created are automatically assigned an accessible URL unless a user manually changes this setting. The company acknowledged this mistake and quickly responded by cutting off access.

The leak was discovered by two security researchers before being confirmed by the American media outlet Fortune. Among the exposed documents, a draft mentions a model named Claude Mythos, also referred to as Capybara. This model represents a new category, surpassing the Opus models, which were previously the most advanced from Anthropic, both in power and cost. Anthropic confirmed to Fortune the authenticity of this project, stating that it is being tested with a select group of clients.

Concerning Offensive Cyber Capabilities

What distinguishes Claude Mythos from previous models goes beyond its reasoning or coding performance. The leaked draft mentions scores significantly higher than those of Opus 4.6 in academic coding and reasoning tests, but it is in the field of cybersecurity that concerns are focused.

The model is described as being well ahead of any other AI model in terms of cyber capabilities, able to exploit software vulnerabilities on a scale that far exceeds the efforts of defenders. In other words, Anthropic admits to having created a tool whose potential offensive capabilities surpass what security teams can counter.

Anthropic is not new to justifying the market release of a model with sensitive capabilities. In February 2026, during the launch of Claude Opus 4.6, the company had already highlighted its model's ability to identify 0-day vulnerabilities in open-source libraries, while acknowledging the double-edged nature of these capabilities.

A Well-Established Communication Strategy

For these reasons, the draft indicates that Claude Mythos will initially be made available to cyber defenders, giving them a head start against an impending wave of AI-driven exploits, as predicted by Anthropic.

This communication fits into a well-established strategy by major players in AI: to highlight capabilities while emphasizing the risks of their new tools, and to present a responsible approach to ease concerns.

In February 2026, OpenAI adopted a similar strategy. During the launch of GPT-5.3-Codex, Sam Altman's company launched a program titled "Trusted Access for Cyber," reserving its most powerful models for defenders through verified access, accompanied by a $10 million fund in API credits for teams working on protecting critical infrastructure.