Anthropic has quietly deployed a new AI model called Claude Mythos Preview as part of Project Glasswing - a cybersecurity partnership with Nvidia, Google, AWS, Apple, Microsoft, and several other large companies.

The model is not available publicly, and Anthropic has no current plans to release it, citing security concerns. Access is restricted to launch partners, who are using it to scan their systems for high-stakes vulnerabilities.

The results so far are notable. Anthropic says Mythos Preview has identified thousands of high-severity vulnerabilities - including some in every major operating system and web browser. Mozilla CTO Bobby Holley confirmed the model found 271 bugs in Firefox 150 alone, describing it as "every bit as capable" as top security researchers.

Crucially, the model did this autonomously. Anthropic's blog post specifically highlights that Mythos Preview identified vulnerabilities and developed related exploits "entirely autonomously, without any human steering."

Though not specifically trained for security, the model's strong agentic coding and reasoning abilities are what Anthropic says drives its effectiveness here. Newton Cheng, Anthropic's cyber red team lead, described the goal as giving defenders a "head start" against adversaries.

The model's existence was first reported in March following a data leak that Anthropic attributed to human error. The official Glasswing announcement came in early April.

Source: The Verge | Anthropic Red Team Blog