The existence of a new AI model, 'Claude Mythos,' was discovered in a draft blog post by Anthropic. Anthropic confirmed the existence of the model and announced that they are testing the new AI model, which they describe as having 'significant performance improvements.'
AI company Anthropic has discovered that it accidentally made public information about its new AI model under development, ' Claude Mythos .' This fact was revealed due to a data leak caused by a misconfiguration of its content management system (CMS). Anthropic has acknowledged the existence of Claude Mythos and explained that it is currently testing it with some customers.
Exclusive: Anthropic 'Mythos' AI model representing 'step change' in power revealed in data leak | Fortune
Exclusive: Anthropic left details of unreleased AI model, exclusive CEO event, in unsecured database | Fortune
https://fortune.com/2026/03/26/anthropic-leaked-unreleased-model-exclusive-event-security-issues-cybersecurity-unsecured-data-store/
The data breach was caused by a human error in the configuration of an external content management system used by Anthropic. Because the assets uploaded to the system were set to public by default, nearly 3,000 assets, including unpublished blog posts, images, and PDFs, were searchable and accessible to anyone.
According to the leaked blog draft, this new AI model is called 'Claude Mythos' and is positioned as a model belonging to a new hierarchy called 'Capybara.' The leaked blog post is archived below.
Claude Mythos - Archive
https://m1astra-mythos.pages.dev/
Capybara is a new model layer that is even larger and more intelligent than the company's previous most powerful Opus model, and is said to 'record dramatically higher scores in software coding, academic reasoning, and cybersecurity testing' compared to the existing Claude Opus 4.6.
Anthropic positions Claude Mythos as a significant step up in performance, but due to high operational costs, it is not yet ready for a general release. As of the time of writing, it is only being tested with a limited number of early access customers.
Anthropic also explained that because Claude Mythos possesses cyber capabilities far superior to previous AI models, they are concerned about the risk of misuse. In fact, there have been reports of hacker groups believed to be linked to the Chinese government attempting to infiltrate approximately 30 organizations, including IT companies and government agencies, using Claude Code, making the management of such a powerful model an urgent issue.
Because Claude Mythos could potentially be exploited faster than existing defensive efforts and there are concerns it could be used by hackers in large-scale cyberattacks, Anthropic is planning a cautious release strategy.
Furthermore, the leaked documents included details about an invitation-only CEO retreat planned to be held at a hotel in the UK for influential European business leaders. Anthropic CEO Dario Amodei was also scheduled to attend the two-day meeting, and participants were expected to experience unannounced features of Claude.
Anthropic has taken steps to make the data private after being pointed out by the overseas media outlet Fortune, stating that 'the leaked content is a draft that is under consideration for public release and will not affect customer data or the security infrastructure.' It should be noted that similar data leaks via websites have occurred in the past at companies such as Apple, Google, and Tesla, and Fortune points out that the proliferation of AI tools is making it easier to find such undisclosed information.
Related Posts:
