Anthropic's new, more powerful AI model confirmed after data leak

Anthropic is testing a new AI model called “Claude Mythos,” the company confirmed after Fortune reported that draft materials about the in-development system had been left in an unprotected, publicly accessible data store on the company’s website.

The company’s spokesperson characterized the new system as “a step change” in AI performance, adding that it was “the most capable we’ve built to date.” The company said it was moving carefully with the rollout because of the model’s power, and that a small group of customers currently has early access.

Related Content

Wikipedia is shutting the door on AI-written articles

Microsoft is joining its peers in mandating a 3-day office week

Fortune reporter Bea Nolan identified the exposed data. Two cybersecurity researchers — Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge — separately assessed the documents at Fortune’s request. Pauwels counted nearly 3,000 files tied to Anthropic’s blog that had not previously appeared on any of the company’s public-facing pages.

Anthropic attributed the exposure to a configuration error in one of its external content management tools, calling it “human error.” After Fortune contacted the company, Anthropic restricted access to the data store.

The draft blog post applied both names — “Claude Mythos” and “Capybara” — to what the document indicated was a single model. According to the document, the system outperformed Claude Opus 4.6 across several benchmarks — including cybersecurity, software coding, and academic reasoning — and would occupy a new tier above Opus in Anthropic’s model lineup. The document also flagged the model’s high running costs and said a public launch had not yet been scheduled.

The documents outlined significant cybersecurity risks associated with the model. The draft blog characterized the system as more advanced in cybersecurity tasks than any competing AI model and warned that it could allow attacks to scale faster than defenders could counter them. Anthropic said the early-access rollout would focus on cyber defense organizations, giving them time to reinforce their systems ahead of a broader release.

The exposed files also contained details of a planned private summit for European business leaders, to be held at a U.K. country manor. Anthropic CEO Dario Amodei is set to attend. Anthropic confirmed the event in a statement, saying it was one in a series of gatherings the company has held for business leaders over the past year.

📬 Sign up for the Daily Brief

Our free, fast and fun briefing on the global economy, delivered every weekday morning.

Sign me up