- A data leak from Anthropic‘s content management system revealed two unreleased AI models codenamed Claude Mythos and Claude Capybara.
- The leaked draft blog post describes the new model as achieving “dramatically higher scores” than Claude Opus 4.6 in coding, reasoning, and cybersecurity benchmarks.
- Anthropic confirmed the model represents “a step change” in AI performance but raised concerns about its cybersecurity capabilities outpacing defenders.
- Security researchers Roy Paz and Alexandre Pauwels discovered roughly 3,000 unpublished assets in an unsecured data cache.
What Happened
Anthropic’s next-generation AI model was revealed not through a planned announcement but through a data leak. Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge discovered approximately 3,000 unpublished assets publicly searchable in an unsecured data cache belonging to Anthropic’s content management system. The exposure was the result of human error in a CMS configuration.
The exposed materials included a draft blog post referencing two model codenames: Claude Mythos and Claude Capybara, which appear to refer to the same model. Fortune first reported the leak on March 26, 2026, after the researchers shared their findings with the publication. Anthropic removed the exposed data on Thursday evening following Fortune’s notification.
Why It Matters
An Anthropic spokesperson confirmed the model’s existence, describing it as representing “a step change” in AI performance and calling it “the most capable we’ve built to date.” If the leaked benchmarks hold up under independent evaluation, Mythos would represent a new tier above Opus 4.6 in Anthropic’s model lineup, potentially repositioning Anthropic at the frontier of AI capabilities.
The disclosure also carries competitive implications. OpenAI, Google DeepMind, and xAI have each released new models in recent months. An unexpected jump in Anthropic’s benchmark performance could shift enterprise and developer adoption patterns, particularly in coding and cybersecurity applications where the leaked draft claims the largest gains.
The leak itself underscores a recurring challenge for AI companies: keeping frontier research under wraps. A misconfigured content management system exposed months of internal planning to the public, forcing Anthropic to acknowledge Mythos before its planned launch timeline.
Technical Details
According to the draft blog post, the new model achieves “dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others” compared to Claude Opus 4.6. The model is currently in early access testing with select customers and is described as expensive to operate, suggesting higher computational requirements than existing Claude models.
Anthropic’s own internal assessment flagged significant cybersecurity concerns. The draft stated the system is “currently far ahead of any other AI model in cyber capabilities” and could enable hackers to “exploit vulnerabilities in ways that far outpace the efforts of defenders.” This dual-use risk appears to be shaping the company’s rollout strategy, with cybersecurity defenders slated for priority access.
The two codenames, Mythos and Capybara, appear in different sections of the leaked materials but reference the same underlying model. Whether one name represents an internal development codename and the other a planned public brand has not been clarified by Anthropic.
Who’s Affected
Anthropic’s existing enterprise customers with early access are the first to evaluate Mythos. The company’s planned rollout prioritizes cybersecurity defenders, aiming to give them access to the model’s capabilities before a broader release. Competing AI labs, including OpenAI and Google DeepMind, now face a potential shift in the frontier model landscape if the claimed benchmark gains are verified.
Researchers Paz and Pauwels, who discovered the leak, contacted Fortune rather than exploiting the data. The roughly 3,000 exposed assets included internal planning documents beyond just the model announcement. The scope of the exposure raises questions about what other proprietary information may have been accessed before Anthropic secured the cache.
What’s Next
Anthropic has not announced a public release date for Mythos. The company’s cautious rollout plan, focused on cybersecurity defenders first, suggests a phased launch designed to mitigate the dual-use risks the company itself identified. The high operational cost mentioned in the draft may also limit initial availability to enterprise customers willing to pay a premium for frontier capabilities.
Whether Anthropic can maintain a controlled rollout after an unplanned public disclosure remains an open question. Independent benchmarking by third-party researchers will be the first test of whether Mythos lives up to the claims in the leaked draft.
