Krux
March 29, 2026
Anthropic's Leaked Claude Mythos Signals Major Capability Jump
Published: March 29, 2026 at 2:51 AM
Updated: March 29, 2026 at 2:51 AM
100-word summary
Anthropic accidentally exposed details of Claude Mythos, its unreleased flagship model, through a misconfigured content management system. The leak revealed a new "Capybara" early-access tier and claims of step-change improvements in reasoning, coding, and cybersecurity tasks. No public release timeline exists yet. The exposure happened through human error, not a security breach, but it highlights a thorny problem: as AI models get better at cybersecurity defense, they also get better at offense. Anthropic is testing Mythos with select customers for secure code generation and threat modeling. The leak suggests the next generation of models will force companies to rethink not just what AI can do, but who gets to use...
What happened
Anthropic accidentally exposed details of Claude Mythos, its unreleased flagship model, through a misconfigured content management system. The leak revealed a new "Capybara" early-access tier and claims of step-change improvements in reasoning, coding, and cybersecurity tasks. No public release timeline exists yet. The exposure happened through human error, not a security breach, but it highlights a thorny problem: as AI models get better at cybersecurity defense, they also get better at offense. Anthropic is testing Mythos with select customers for secure code generation and threat modeling.
Why it matters
The leak suggests the next generation of models will force companies to rethink not just what AI can do, but who gets to use it and when.