Anthropic’s Secret ‘Capybara’ Model Leaked: So Powerful Even the Company Hesitates to Release It

Anthropic accidentally exposed internal draft of its next‑generation model, internally codenamed Capybara and marketed as Mythos , after a CMS configuration error left thousands of assets publicly accessible.

"Capybara currently leads other AI models by a wide margin in cybersecurity capability. It heralds a new wave of models that can discover and exploit software vulnerabilities faster than defenders can patch them."

The leaked document states the model surpasses the top‑of‑the‑line Claude Opus 4.6 in software programming, academic reasoning, and cybersecurity tests, and is larger, smarter, and more expensive than the existing Opus tier.

Anthropic’s spokesperson confirmed to Fortune that this represents a “step‑change in capability” and that only a handful of early‑access customers have tested it.

The leak originated on March 26 when Fortune reporters discovered Anthropic’s content‑management system defaulted to public access. Cambridge researcher Alexandre Pauwels, commissioned by Fortune, found nearly 3,000 assets—including draft blog posts, images, and PDFs—exposed.

Anthropic responded by revoking public access, labeling the incident a human error in CMS configuration and emphasizing it was unrelated to Claude or AI code.

Beyond the security concerns, the article notes rumored model size of 100 trillion parameters and $10 billion training cost, positioning Capybara above competitors such as OpenAI’s GPT‑5.4 and Google’s Gemini 2 Ultra.

Anthropic plans to first grant access to cybersecurity partners to allow defenders to prepare before broader release; the official launch timeline remains undefined and will depend on defensive readiness and competitive pressure.