Anthropic has developed a new AI model called “Claude Mythos Preview,” but the company has decided not to release it publicly due to serious security concerns. The model proved highly effective at identifying and exploiting critical vulnerabilities in operating systems and web browsers, uncovering thousands of previously unknown security flaws that had gone undetected for years.

Instead of making Mythos widely available, Anthropic is providing access through its tightly controlled security initiative, “Project Glasswing.” In collaboration with other major tech companies, the model is intended for defensive use to help protect critical software infrastructure. This cautious approach highlights both the potential and the risks of advanced AI systems.