On March 26, Fortune reported that security researchers discovered roughly 3,000 unpublished assets in a publicly accessible data cache due to a CMS misconfiguration at Anthropic. Among them: a draft blog post revealing a new model tier codenamed 'Capybara,' marketed externally as 'Claude Mythos.'
The leaked draft describes Mythos as larger and more intelligent than Opus, calling it 'a step change' with 'dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity.' Anthropic's own draft warned that the model poses 'unprecedented cybersecurity risks' and is 'currently far ahead of any other AI model in cyber capabilities.'
The rollout reportedly starts with early-access customers specifically for cybersecurity evaluation before broader availability. This suggests Anthropic is taking a cautious, staged deployment approach — letting security researchers probe the model before general release.
Anthropic has not officially confirmed or denied the leak. The cybersecurity warning is notable — it's rare for an AI company to flag its own model as a potential risk in marketing materials. If real, Mythos would represent a new tier above Opus in Anthropic's lineup, potentially reserved for enterprise and research customers given the cost and safety implications.