A data leak at Anthropic has revealed that it is testing a powerful AI model known as Claude Mythos, or Capybara. It is so powerful that developers are concerned about the speed of cyberattacks that can be carried out using Mythos.
A data leak at Anthropic has revealed that it is testing a powerful AI model known as Claude Mythos, or Capybara. It is so powerful that developers are concerned about the speed of cyberattacks that can be carried out using Mythos.
This became known after working materials on the model were left in an unsecured, publicly accessible data repository on the Anthropic website.
Two cybersecurity researchers—Roy Paz of LayerX Security and Alexander Pauwels of the University of Cambridge—independently analyzed the documents at the request of Fortune. Pauwels counted nearly 3,000 files related to the Anthropic blog that had not previously appeared on any of the company’s public pages.
Anthropic attributed the data leak to a configuration error in one of its external content management tools, calling it «human error.» The company subsequently restricted access to the data repository.
In a draft blog post, the names Claude Mythos and Capybara referred to a single model that outperformed Opus 4.6 on several metrics, including cybersecurity, software coding, and academic reasoning. The paper also noted the model’s high operating costs and said that a public launch date had not yet been determined. The developers also believe it poses unprecedented cybersecurity risks.
An Anthropic representative called the new model a «major step» in AI development and «the most powerful we’ve created to date.» The company said the model is currently being tested by «early access users.»
«We are developing a general-purpose model with significant advances in logical reasoning, coding, and cybersecurity. Given the power of its capabilities, we are carefully considering how to release it. As is customary in the industry, we are working with a small group of early access customers to test the model. We consider this model to be a breakthrough and the most powerful one we have created to date,» said an Anthropic representative.
The leaks highlighted significant cybersecurity risks associated with the model. A draft blog post described the system as more advanced at cybersecurity tasks than any competing AI model, and included a warning that it could allow attacks to spread faster than defenders could counter them.
Anthropic said the early access rollout will focus on cybersecurity organizations, giving them time to harden their systems before a broader release.
