Claude Mythos: Anthropic's AI Held Back Over Cyber Risks

Anthropic developed an AI model called Claude Mythos that was held back from public release due to cybersecurity risks. This model, introduced in April 2026, is Anthropic's most advanced for reasoning, software engineering, and cybersecurity tasks. Internal testing showed Claude Mythos could identify software vulnerabilities and develop exploit paths with minimal human guidance. It achieved a perfect score on the Cybench cybersecurity benchmark and found thousands of unknown software flaws, including a 17-year-old vulnerability in FreeBSD. Anthropic worried that if widely released, attackers could use Mythos to exploit weaknesses faster than they could be fixed. Because of this, the company restricted access and launched Project Glasswing. This initiative uses Mythos for defensive cybersecurity purposes with major tech and security organizations. The bottom line is that advanced AI capabilities can present complex safety challenges, leading companies to prioritize responsible deployment over immediate public access.