Anthropic's restricted Claude Mythos AI leaked to unauthorized users before official launch, undermining its 'safety-first' brand

Anthropic announced Claude Mythos, a new AI model it claimed was too capable at hacking and cybersecurity to release publicly, and began controlled testing with select companies. Within days of that announcement, Bloomberg reports unauthorized users gained access to the model anyway — a breach Anthropic is now investigating.

The company built its reputation on prioritizing AI safety over speed, positioning itself as more cautious than competitors like OpenAI. This leak directly contradicts that core brand promise, showing a gap between Anthropic's stated security practices and what actually happened.

For business leaders evaluating AI vendors, this signals risk: a company trusted with sensitive AI access failed to control that access from day one. For employees at companies in Anthropic's testing program, it raises questions about whether their proprietary data or use cases were exposed to the same unauthorized group.