Anthropic Confirms Testing of Claude Mythos, After Draft Announcement Leaked Online
Anthropic has confirmed it is testing Claude Mythos, a new general-purpose AI model the company describes internally as by far the most powerful it has ever developed.
The existence of the model became public on March 26 after a configuration error in the company's content management system left a draft blog post and nearly 3,000 unpublished assets exposed in a publicly searchable data cache.
The leak, first reported by Fortune, included structured web-page data outlining the planned announcement of Claude Mythos, along with images, PDFs, and other files.
Anthropic quickly restricted access to the cache after being contacted by the publication and attributed the exposure to human error in how its external CMS tool defaults new assets to public visibility.
A few cybersecurity researchers independently discovered the materials before the company locked them down.
In a statement provided to Fortune, an Anthropic spokesperson said the company had completed training on the model and was proceeding with a measured rollout.
“We’re developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity,” the spokesperson said. “Given the strength of its capabilities, we’re being deliberate about how we release it. As is standard practice across the industry, we’re working with a small group of early access customers to test the model. We consider this model a step change and the most capable we’ve built to date.”
The draft blog post reviewed in the leak referred to the all-new AI model as Claude Mythos and positioned it as the debut of a new performance tier internally called Capybara.
This tier sits above the current Opus family of models, which until now represented Anthropic's flagship offerings. The document noted that Mythos delivers dramatically higher scores than Claude Opus 4.6 on benchmarks covering software coding, academic reasoning, and cybersecurity tasks.
Training is finished, but the model remains expensive to operate and is not yet prepared for broad availability.
Anthropic highlighted particular concerns around cybersecurity in the leaked materials.
The draft stated that Mythos is currently far ahead of any other AI model in cyber capabilities and “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”
The company plans to prioritize early access for organizations focused on cyber defense, giving them time to strengthen codebases before wider exposure to potential AI-driven exploits.
This approach aligns with the firm's broader safety-focused philosophy, which has shaped previous model releases.
The timing of the leak comes roughly two months after Anthropic publicly launched Claude Opus 4.6 on February 5 as its then-most advanced system for coding, agentic workflows, and professional tasks.
No official blog post or product page for Mythos appears on Anthropic's website as of March 30, and the company has not disclosed a public release schedule.
Industry observers note that frontier AI labs routinely conduct closed-door testing of high-capability systems before broader deployment, especially when dual-use risks such as offensive cyber applications are involved.
Claude Mythos remains in limited early-access testing with selected customers, and Anthropic has given no indication of when or how it will become available beyond that initial group.
The company's cautious stance underscores the growing intersection of advanced AI capabilities and real-world security implications at the current frontier.
