Anthropic's New Model Claude Mythos Accidentally Leaked: Cybersecurity Capabilities Far Surpass Existing AI

Anthropic's next-generation model Claude Mythos was accidentally leaked due to a content management system configuration error, with unreleased blog drafts stored in a publicly searchable database. Anthropic confirmed the model as its strongest to date, representing a step change in capability.

Claude Mythos is positioned above the Opus series, with significantly higher scores in software programming, academic reasoning, and cybersecurity tests than Claude Opus4.6. Its cybersecurity capabilities far surpass all existing AI models. Anthropic is adopting a cautious release strategy, first opening it to a small number of early customers for cybersecurity defense and not making it public in the short term.

The leak stemmed from human error in an external CMS tool, exposing nearly 3,000 unreleased assets. Researchers from the University of Cambridge and LayerX Security independently discovered the leak, which also included details of Anthropic's closed-door European corporate CEO summit in the UK.