Anthropic confirmed it’s testing Claude Mythos after someone spotted an unpublished blog post in a public data store. The company called it a human error in their content management system that exposed around 3,000 draft assets.
Fortune first reported the leak, and Anthropic told them the model marks “a step change” and stands as “the most capable we’ve built to date.” According to Technobezz, they’re sharing early access with a small group of customers to check its behavior and risks, with no broad release on the horizon.
What leaked about Claude Mythos
The draft post described Mythos as kicking off a new model tier named Capybara, which goes beyond their current Opus lineup. It crushes Opus 4.6 on benchmarks for software coding, academic reasoning, and cybersecurity tasks.
Anthropic’s own docs warn Mythos leads all other AI models in cyber skills right now. They see it finding and exploiting vulnerabilities faster than defenders can keep up, so they’re prioritizing access for security teams first. Details from Mint and Times of India match up here.
Why the hold-up
Besides the cyber worries, Mythos eats a ton of compute, making it pricey to run. Anthropic wants to tweak efficiency before opening it wider. A spokesperson emphasized advances in reasoning, coding, and cybersecurity, but they’re moving slow on purpose.
The leak also spilled details on an upcoming executive retreat in Europe for AI demos, but that’s secondary to the model news.
