Anthropic is developing a new AI model, reportedly more capable than its previous releases. The existence and details of this model were inadvertently revealed following a data leak.
An Anthropic spokesperson described the model as a "step change" in AI performance and the "most capable" built to date, confirming it is currently being trialed by early access customers.
Data Leak Reveals Anthropic's Next-Gen AI
Descriptions of the model, including a draft blog post, were stored in a publicly accessible data cache and reviewed by Fortune. This cache contained close to 3,000 assets linked to Anthropic's blog that were not previously published. The material was in an unsecured, public data lake due to what Anthropic later acknowledged as a "human error" in the configuration of its content management system (CMS).
After being informed of the leak, Anthropic removed public access to the data store. The company stated that the accessible material consisted of "early drafts of content considered for publication."
Introducing "Claude Mythos" and "Capybara"
The leaked draft blog post referred to the new model as "Claude Mythos" and also introduced "Capybara" as a new tier of AI model.
"Capybara" is described as larger, more intelligent, and more expensive than the company's previous most powerful Opus models. It showed significantly higher scores in tests of software coding, academic reasoning, and cybersecurity compared to Claude Opus 4.6. "Claude Mythos" was described as "by far the most powerful AI model we've ever developed."
Anthropic confirmed it is developing a general-purpose model with advances in reasoning, coding, and cybersecurity. It is working with a small group of early access customers for testing due to the model's advanced capabilities.
Cybersecurity Concerns and Strategic Release
The leaked document indicated that the new AI model poses significant cybersecurity risks.
Anthropic expressed concern that the system is "currently far ahead of any other AI model in cyber capabilities" and could enable new types of AI-driven exploits. The company's plan for the model's release focuses on cyber defenders, offering early access to organizations to help them improve the robustness of their codebases against potential AI-driven attacks.
Anthropic and OpenAI have noted similar cybersecurity risks with their latest frontier models. OpenAI classified its GPT-5.3-Codex as "high capability" for cybersecurity tasks, and Anthropic's Opus 4.6 also demonstrated dual-use capability in identifying software vulnerabilities. Anthropic has also reported instances of state-sponsored hacking groups attempting to exploit Claude in cyberattacks.
Executive Retreat Details Also Revealed
Among the leaked documents was a PDF containing information about an upcoming, invite-only retreat for European CEOs in the U.K., which Anthropic CEO Dario Amodei will attend.
The two-day event, described as an "intimate gathering" at an 18th-century manor, is intended for attendees to discuss AI adoption and experience unreleased Claude capabilities.
Anthropic confirmed the event as part of an ongoing series.