Anthropic has confirmed a security breach involving its Claude Mythos Preview AI model, which autonomously identifies zero-day vulnerabilities. Unauthorized access was gained through compromised contractor credentials and data from a separate breach at Mercor, an AI training data provider. The breach occurred just two weeks after Mythos was announced on April 7, 2026. In response, Anthropic launched Project Glasswing, a $100 million initiative to allow vetted organizations to use Mythos for defensive cybersecurity purposes. The breach highlights vulnerabilities in third-party vendor security and raises concerns about AI governance and regulatory risks. The incident underscores the need for robust access controls and vendor audits in the AI industry.