A security misconfiguration has exposed Anthropic's next-generation AI model before its official announcement, revealing a system so powerful that its creators warn it could accelerate cyber warfare.
The AI company confirmed it is testing "Claude Mythos," which it describes as "the most capable we've built to date" and represents "a step change" in performance. Early access customers are currently evaluating the model while Anthropic assesses its behavior and potential risks.
Approximately 3,000 unpublished assets linked to Anthropic's blog became publicly accessible through what the company called a "human error" in its content management system configuration. The exposed materials included draft blog posts and internal documents that had not been cleared for public release.
According to leaked documents reviewed by Fortune, Claude Mythos introduces an entirely new tier of AI models called "Capybara" that sits above Anthropic's current Opus-class offerings.
"larger and more intelligent than our Opus models, which were, until now, our most powerful."
What makes this leak particularly significant is the model's unprecedented cybersecurity capabilities. Internal warnings state Mythos is "currently far ahead of any other AI model in cyber capabilities" and "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
Anthropic appears especially concerned about dual-use applications where the same technology that helps security professionals find vulnerabilities could also empower malicious actors.
The company plans a cautious rollout focused initially on cyber defenders, giving organizations "a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits."
News of the security incident triggered market reactions across technology sectors. Cybersecurity stocks including Palo Alto Networks, Crowdstrike and Fortinet dropped 4-6% following news of the leak and the model's advanced capabilities. The broader iShares Expanded Tech-Software Sector ETF declined 2.5%.
Beyond technical specifications, the data breach also revealed plans for an exclusive executive retreat in Europe scheduled for later this year. The invite-only gathering at an 18th-century English manor will host European CEOs and feature discussions about AI adoption alongside demonstrations of unreleased Claude capabilities.
Anthropic acknowledged the configuration error led to draft content being accessible but emphasized it has implemented corrective measures to prevent further unauthorized access.
Company representatives described ongoing development of a "general-purpose model with meaningful advances in reasoning, coding, and cybersecurity."
Separate reporting indicates Anthropic is evaluating a potential public offering targeted for October 2026, which would position the company alongside OpenAI's anticipated market debut.
