Anthropic's Secret Claude Mythos Model Revealed in Data Leak - Outperforms Opus on Every Benchmark

Anthropic accidentally leaked details of its next major AI model, Claude Mythos, through an unsecured public data store. A draft blog post and roughly three thousand unpublished assets were left in an unencrypted cache - and someone grabbed the blog post before it was pulled down. According to the leaked documents, Mythos represents what Anthropic internally calls a step change in capabilities. Benchmarks show it dramatically outperforming Claude Opus 4.6 across software coding, academic reasoning, and cybersecurity tasks. The model is reportedly already being trialled by early access customers under the internal codename Capybara. The leak also flagged serious security concerns. Internal assessments warn that Mythos could rapidly find and exploit software vulnerabilities at a scale that would accelerate the cyber arms race. Anthropic's own safety teams flagged the cybersecurity risk as unprecedented for any model they've tested. The data exposure was independently discovered by researchers at Cambridge and LayerX Security, which suggests the cache was sitting in the open long enough for multiple parties to find it. An accidental leak revealing a model your own safety team says poses unprecedented cyber risk - that's the kind of story that writes itself. Anthropic will need to address both the capability claims and the fact that their own infrastructure let it slip.