Anthropic's Claude Mythos Leaked: Most Powerful AI Model Yet Exposed

Anthropic, the AI company behind the popular Claude chatbot, has accidentally revealed its most powerful artificial intelligence model to date through a significant data leak. The company confirmed it is testing "Claude Mythos," describing the new system as a "step change" in AI capabilities that represents their most advanced technology ever developed. According to Fortune, the leaked documents were discovered by cybersecurity researchers from LayerX Security and the University of Cambridge.

The Accidental Revelation

The leak occurred when draft blog posts and internal documents were inadvertently stored in a publicly accessible data lake. In total, nearly 3,000 unpublished assets linked to Anthropic's blog were found to be publicly accessible due to a configuration error in the company's content management system.

According to the leaked documents reviewed by cybersecurity experts, the new model called "Claude Mythos" is described as "by far the most powerful AI model we've ever developed." The documents also referenced a new tier of AI models named "Capybara," which Anthropic says is "larger and more intelligent than our Opus models" that were previously their most capable offerings.

An Anthropic spokesperson confirmed the leak in a statement, acknowledging that "human error" in their CMS configuration led to the exposure. The company has since secured the data store and removed public access to the draft materials.

Unprecedented Capabilities and Cybersecurity Risks

The leaked documents reveal that Claude Mythos achieves dramatically higher scores than previous models on tests of software coding, academic reasoning, and cybersecurity tasks. However, Anthropic's own internal assessments highlight significant concerns about the model's potential misuse.

According to the leaked blog post, the model is "currently far ahead of any other AI model in cyber capabilities," and Anthropic warns that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." The company explicitly states that hackers could potentially use the model to execute large-scale cyberattacks.

This dual-use nature of advanced AI capabilities has become a growing concern across the industry. In February 2026, OpenAI similarly warned that its GPT-5.3-Codex model was the first it classified as "high capability" for cybersecurity-related tasks under its Preparedness Framework. Anthropic's Opus 4.6, released the same week, also demonstrated abilities to surface previously unknown vulnerabilities in production codebases.

The company has previously documented real-world attempts by hacking groups, including those linked to the Chinese government, to exploit Claude for cyberattacks. In one case, Anthropic discovered a state-sponsored group using Claude Code to infiltrate approximately 30 organizations before the company detected and stopped the operation.

Cautious Rollout Strategy

Anthropic appears to be taking a more measured approach to releasing Claude Mythos compared to previous launches. According to the leaked documents, the company is implementing a cautious rollout focused specifically on cybersecurity defenders rather than general public release.

"We're releasing it in early access to organizations, giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits," the draft blog post stated. The model is currently being tested with a small group of early access customers, and Anthropic notes it is "expensive to run and not yet ready for general release."

The new Capybara tier represents a shift in Anthropic's product lineup. Currently, the company markets models in three sizes: Opus (largest and most capable), Sonnet (balanced performance), and Haiku (fastest and most affordable). Capybara establishes a new premium tier above Opus, offering enhanced capabilities at a higher computational cost.

This strategic approach reflects the growing tension in the AI industry between rapid innovation and responsible deployment. As models become increasingly powerful, companies like Anthropic are grappling with how to make advanced capabilities available while mitigating potential harms.

The data leak also revealed plans for an exclusive CEO retreat in the English countryside, where Anthropic CEO Dario Amodei and European business leaders will discuss AI adoption and experience unreleased Claude capabilities. The company described this as part of an ongoing series of events aimed at engaging corporate customers.

As the AI arms race intensifies between major players including OpenAI, Google, and Anthropic, the emergence of Claude Mythos signals that the next generation of AI models may bring capabilities that challenge existing security frameworks. The incident serves as a reminder that even AI safety-focused companies face significant challenges in both securing their own systems and responsibly managing increasingly powerful technologies.