Anthropic has limited access to its unreleased Claude Mythos model following internal tests demonstrating advanced cybersecurity capabilities, including the identification and exploitation of thousands of software vulnerabilities across major systems.