Anthropic Accidentally Leaked a Model They Call an 'Unprecedented Threat.'
On March 27th, roughly three thousand unpublished Anthropic documents appeared online - including a draft blog post describing Claude Mythos, a brand new AI model in the Capybara class. Stronger than anything the company has ever built.
The cause? A website configuration error. Files set to public by default. The discovery was made independently by Roy Paz from LayerX Security and Alexandre Pauwels from the University of Cambridge. Fortune broke the story first.
Anthropic confirmed the model exists. They called it “a step change” and “the most powerful model we’ve ever built.”
What is Capybara
To understand the scale - Anthropic has three model classes. Haiku - fastest and cheapest. Sonnet - middle tier. Opus - most powerful.
Capybara is a new, fourth class. Above Opus. Stronger than anything Anthropic has had before. Claude Mythos is the first model in this class.
According to the leaked blog draft, Mythos significantly outperforms Claude Opus 4.6 in coding, complex problem-solving, and cybersecurity.
”Unprecedented risk”
Here’s the unsettling part. In the leaked documents, Anthropic itself warns that the model poses “unprecedented cybersecurity risk” - the ability to rapidly find and exploit software vulnerabilities at a level that surpasses anything that existed before.
That’s why there won’t be a public release. For now, the model is being tested by a closed group of early customers, with particular focus on the cyber defense sector.
Markets reacted
Cybersecurity stocks - Palo Alto Networks, CrowdStrike, Fortinet - dropped 4 to 6 percent on the news. Bloomberg and The Information report that Anthropic is considering an IPO as early as October this year.
The leaks also included details of a closed-door CEO summit at an 18th-century English estate, attended by Dario Amodei.
My take
If this holds up, Anthropic is extending its lead over OpenAI and Google. But at the same time, it’s admitting its own tools are becoming dangerous.
A company building a model it considers a threat - while planning to go public in the same quarter. Responsible safety approach? Or IPO hype? Probably both at once.
This is part of AITU #02 - a weekly roundup of AI and tech news. The full episode with 7 stories from March 20-27, 2026 is available on my YouTube channel.