Anthropic data leak reveals secret AI model 'Mythos' — allegedly most powerful ever
What it really says
Two security researchers found approximately 3,000 internal Anthropic files publicly accessible — including a blog post draft about a model called 'Mythos'. According to the draft, it's 'by far the most powerful AI model we've ever developed' and can autonomously execute action sequences across multiple systems. Anthropic warned US authorities that Mythos could make large-scale cyberattacks more likely.
Our assessment
Worth looking at closely: First, this was a data leak, not an intentional release — showing that even 'safety-focused' companies make mistakes. Second, the warning to US authorities is actually a good sign: the company takes risks seriously enough to involve the government. Third: 'autonomously across multiple systems' sounds dramatic but is a logical evolution of existing AI agents.
Relevance for Germany
For German companies, the question arises: which AI providers do we trust? Anthropic positions itself as safety-conscious — but this leak undermines exactly that trust.
Fact check
The data leak was confirmed by two independent security researchers. Anthropic confirmed Mythos' existence after the leak and stated the model is in testing phase.
Source
- • Fortune Reporting 26.03.2026
- • The New Stack 27.03.2026
- • Anthropic Official Statement