A configuration error at Anthropic has inadvertently revealed the company's most secret technological advantage. According to an exclusive report from Fortune on the 26th, a new flagship model named "Claude Mythos," with the internal codename "Capybara," has completed training and entered early testing phases. Anthropic officially describes its performance improvement as a "generational leap," stating it comprehensively surpasses the current top-tier version, Claude Opus 4.6, in core tests such as programming, academic reasoning, and cybersecurity.

The report indicates that due to a human configuration error in Anthropic's content management system (CMS), nearly 3,000 unreleased internal documents were exposed in a publicly searchable data cache. The leaked files reveal the company plans to introduce a new model tier named "Capybara," positioned above the existing flagship Opus series, with correspondingly higher operational costs.

However, this superior performance comes with unprecedented security concerns. Leaked internal drafts classify the new model as a significant risk source in the cybersecurity domain, suggesting its capabilities for cyber attacks "far exceed those of any other current AI model." The drafts warn that if exploited by malicious actors, it could enable large-scale cyber attacks whose destructive potential would likely outpace defenders' capabilities—a core reason for the company's hesitation to release it publicly.

**Surpassing Opus: New "Capybara" Tier Disrupts the Status Quo** The leaked blog drafts indicate that Claude Mythos represents a structural reorganization of Anthropic's product lineup. Currently, Anthropic's model portfolio consists of three tiers: the most capable Opus, the speed-and-cost-balanced Sonnet, and the smallest and fastest Haiku. The leaked documents suggest Anthropic is introducing a new "Capybara" tier, positioned above Opus—larger in scale, higher in intelligence, but also more expensive to operate. Regarding specific performance, the draft blog states that "compared to the previous strongest version, Claude Opus 4.6, Capybara shows significant score improvements in tests for software programming, academic reasoning, and cybersecurity." The files describe Claude Mythos as "the most powerful AI model we have ever developed, far surpassing any previous version."

An Anthropic spokesperson confirmed the general direction in response to Fortune's inquiry, stating the company is "developing a general model that has made significant advances in reasoning, programming, and cybersecurity," and emphasizing that "given the strength of its capabilities, we are proceeding carefully with its release strategy."

**Unprecedented Cybersecurity Risks: Anthropic Itself Expresses Caution** Accompanying the powerful capabilities are unusually strong security warnings within the leaked files. The draft blog explicitly states the new model "currently far surpasses any other AI model in capabilities for cyber attacks" and signals "an incoming wave of models whose speed of vulnerability exploitation will far outpace defenders' ability to respond." Precisely due to these concerns, Anthropic's release strategy prioritizes cybersecurity defense organizations as the primary beneficiaries. The draft notes: "We will grant early access to relevant institutions first, allowing them to get a head start in fortifying the robustness of their codebases against the impending wave of AI-driven vulnerability exploits."

This concern is not isolated. In February, when OpenAI released GPT-5.3-Codex, it was listed as the first model to achieve a "high capability" rating in cybersecurity tasks and the first directly trained to identify software vulnerabilities. Around the same time, Anthropic's release of Claude Opus 4.6 also demonstrated an ability to identify unknown vulnerabilities in production codebases, with the company acknowledging this as a "double-edged sword" characteristic applicable to both offense and defense.

**Configuration Error: 3,000 Internal Files Accidentally Made Public** The technical root of this leak lies in what appears to be a basic operational mistake. Roy Paz, a senior AI security researcher at LayerX Security, and Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge, discovered a configuration flaw in the external CMS tool Anthropic uses for its official blog: the system defaulted to setting uploaded digital assets as public with accessible URLs unless manually set to private by the user. This default setting led to nearly 3,000 unpublished assets—including images, PDFs, and audio files—being exposed in a publicly searchable data cache.

Anthropic attributed the incident in a statement to "human error," stating "an issue with an external CMS tool led to draft content being publicly accessible," and characterized the leaked materials as "early drafts of content under consideration for release."

**X Platform Erupts: A Mix of Astonishment and Skepticism** Once the news broke, the X platform quickly became a hub of discussion, with public focus split between technological awe and security trust. An account named TFTC highlighted the irony of the situation: "Anthropic inadvertently proved the AI safety argument. A CMS config error led to 3,000 unpublished files being exposed in a publicly searchable cache—including details about 'Claude Mythos' (codename Capybara), which Anthropic's internal files call 'the most powerful AI to date.'" This comment spread widely within tech circles, with many users expressing astonishment that a company with AI safety as a core value had exposed its own hand due to a basic configuration error.

Another user, fardeen, shifted perspective to Claude's latest capabilities, commenting: "Claude can now actually use a computer like you do—open apps, click buttons, fill out forms. Anthropic is gradually removing humans from the operational loop."

User Oliwier Makowski Trusz suggested the reveal of Capybara changes the competitive landscape. Based on the leaked information, the model has reached a parameter count of 10 trillion. The gap between Claude and all other models has just widened significantly.