Anthropic accidentally leaked its next-generation model "Claude Mythos": a new level beyond Opus, with cybersecurity capabilities "far surpassing all existing AIs."

According to monitoring by 1M AI News, Anthropic is testing a new model called “Claude Mythos,” which was accidentally exposed due to a misconfiguration in the content management system that stored an unpublished blog draft in a publicly searchable database. Anthropic subsequently confirmed the existence of this model, stating that it represents a “step change” in capabilities and is “the strongest model we have built to date.”

The leaked draft indicates that Mythos is a completely new model tier, positioned above the Opus series. The draft claims that “compared to the previously best model Claude Opus 4.6, this model shows significant improvements in software programming, academic reasoning, and cybersecurity tests.” The draft also mentions the codename “Capybara,” which points to the same underlying model as Mythos.

Anthropic is particularly cautious about the cybersecurity capabilities of this model, with the draft stating that it “currently far outpaces any other AI model in terms of cybersecurity capabilities” and “foretells an upcoming wave of models whose exploitation abilities will far exceed defenders’ response speeds.” As a result, Anthropic has adopted a slower release strategy than before: initially opening it up to a small number of early customers, focusing on cybersecurity defense scenarios, helping organizations strengthen their codebases before the onslaught of AI-driven attacks. The operational costs of this model are extremely high, and it will not be made publicly available in the short term.

The leak originated from a “human error” in Anthropic’s external CMS tool, leading to nearly 3,000 unpublished assets (including blog drafts, images, and PDFs) being publicly stored in an unencrypted database. Cambridge University cybersecurity researcher Alexandre Pauwels and LayerX Security senior AI security researcher Roy Paz independently discovered these files. The leaked content also includes details of a closed-door summit for European CEOs that Anthropic is set to host in the UK, which CEO Dario Amodei will attend, located at an 18th-century manor hotel in the English countryside.

An Anthropic spokesperson stated: “We are developing a general model that has made significant progress in reasoning, programming, and cybersecurity. Given the strength of its capabilities, we will be very cautious in our release approach.”