- Anthropic confirmed it is developing Claude Mythos (internal codename Capybara), a model it describes as “a step change and the most capable we’ve built to date,” after security researchers discovered nearly 3,000 internal files in an unsecured, publicly accessible data store.
- A leaked draft blog post described Mythos as “a new tier of model: larger and more intelligent than our Opus models” with dramatically higher scores on coding, academic reasoning, and cybersecurity benchmarks.
- The company acknowledged the model is “currently far ahead of any other AI model in cyber capabilities,” raising concerns that it could enable vulnerability exploitation faster than defenders can respond.
- Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge discovered the exposed data store.
What Happened
On March 27, 2026, Fortune reported that a misconfigured content management system at Anthropic left approximately 3,000 internal files publicly accessible without authentication. Among the exposed materials was a draft blog post announcing Claude Mythos, internally codenamed Capybara, a new AI model positioned above Anthropic’s existing Opus line.
Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge discovered the unsecured data store and notified Fortune, which reviewed the materials before publishing its report. Anthropic removed public access to the exposed files later that Thursday evening. The leaked files also included employee parental leave records, PDFs about a planned CEO summit in the UK, images, and audio files.
Why It Matters
The leak is significant on two fronts. First, it revealed a model that Anthropic had not publicly announced, forcing the company into a premature confirmation. An Anthropic spokesperson acknowledged the company is “developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity” and called it “a step change and the most capable we’ve built to date.”
Second, the security failure itself is damaging for a company that has built its brand around AI safety and responsible deployment. Leaving 3,000 internal documents in an unencrypted, publicly searchable database contradicts the careful operational security image Anthropic cultivates. The company has repeatedly emphasized its commitment to safety protocols and responsible scaling policies, making an unforced data exposure particularly embarrassing.
The incident also raises questions about internal data handling at AI labs more broadly. If a company as safety-focused as Anthropic can misconfigure a public data store, the risk of similar incidents at less security-conscious organizations is considerable.
Technical Details
The leaked draft blog post described Mythos as “a new tier of model: larger and more intelligent than our Opus models — which were, until now, our most powerful.” Compared to Claude Opus 4.6, Capybara achieves “dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others.” The positioning above the Opus tier suggests a model with significantly more parameters and training compute than Anthropic’s current flagship.
The cybersecurity capabilities drew particular attention. The draft stated the model 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.” This language suggests Anthropic’s internal testing found the model capable of autonomously identifying and exploiting software vulnerabilities at a level that exceeds current defensive tools.
Anthropic said the model is currently being trialed by “early access customers.” The root cause of the exposure was human error in content management system configuration, not a targeted attack or sophisticated breach.
Who’s Affected
Anthropic’s competitors, particularly OpenAI and Google DeepMind, now have advance visibility into Anthropic’s next-generation capabilities and roadmap. Early access customers testing Mythos face uncertainty about whether the leak compromises any proprietary integration work or confidential evaluation results.
The cybersecurity community has raised alarms about the model’s described offensive capabilities. If Mythos can exploit vulnerabilities faster than defenders patch them, its release timeline and access controls become matters of public interest beyond the AI industry. Government regulators tracking AI risk may also take note, as models with autonomous cyber-offensive capabilities fall squarely within emerging AI safety frameworks in the US, EU, and UK.
Anthropic employees whose personal records were among the exposed files face potential privacy consequences, adding a human dimension to what is otherwise a corporate and technical story.
What’s Next
Anthropic has not announced a public release date for Mythos. The company’s spokesperson confirmed the model exists and is in early testing, but provided no timeline. The leak may accelerate Anthropic’s disclosure plans, as withholding details about a model already described in leaked documents offers diminishing strategic value. How Anthropic addresses the security lapse itself, and whether additional internal materials surface from the window of public access, will test the company’s crisis response.
