Anthropic Withholds Mythos Model Over Security Risks

Anthropic announced on April 12, 2026, that it will withhold the public release of its new Claude Mythos model ^[1]. The company stated the frontier AI is too powerful for general availability ^[1][2]. Mythos can reportedly identify and exploit thousands of high-severity vulnerabilities in operating systems and web browsers ^[1][3]. To manage these risks, the company launched Project Glasswing ^[1]. This initiative allows select security partners to test the model in a private, controlled setting ^[1][2]. The goal is to identify and patch critical software flaws before the technology can be misused ^[2]. Some industry experts view the move as a strategic bid to lead in the AI safety sector ^[3]. While the model remains private, Anthropic continues to emphasize the importance of rigorous security evaluations ^[2][3]. The company did not specify when or if a public version would arrive ^[1].