Exclusive: Anthropic 'Mythos' AI model representing 'step change' in power revealed in data leak

Key Points:

• Anthropic is developing a new AI model called Claude Mythos (also referred to as Capybara), which it describes as the most capable and powerful model it has built to date, currently being trialed by early access customers.
• A data leak occurred due to a human error in the configuration of Anthropic’s content management system, exposing nearly 3,000 unpublished assets including draft blog posts detailing the new model and its significant cybersecurity risks.
• The leaked documents reveal Anthropic’s concerns that Claude Mythos poses unprecedented cybersecurity risks, with capabilities far exceeding previous models, potentially enabling large-scale AI-driven cyberattacks.
• Anthropic plans a cautious rollout focusing on cybersecurity defenders to help them prepare for the wave of AI-driven exploits, highlighting the dual-use nature of such advanced AI in both offense and defense.
• The leak also exposed details of an upcoming invite-only CEO retreat in the UK, where European business leaders will discuss AI adoption and preview unreleased Claude capabilities, part of Anthropic’s strategy to engage large corporate customers.