Anthropic launches Mythos, a new cybersecurity model so powerful it is tightly restricted

Anthropic has introduced Mythos, a new frontier AI model focused on cybersecurity, as part of its Project Glasswing initiative. The company describes it as a “step change” in capabilities, and while the preview is not generally available due to security concerns, it is being offered to 12 partners, including Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks, along with 40 additional organizations that are using the model to help find and fix vulnerabilities across Anthropic’s own systems and open source software.

Anthropic says the model has already found thousands of high severity zero day vulnerabilities in recent weeks, with many of the flaws dating back more than a decade. The company says that although Mythos was not specifically trained for cybersecurity, its coding and reasoning capabilities make it effective for real world software analysis, while internally it has been described as more capable than the Opus line in areas such as coding, reasoning, and academic tasks. Ironically, leaked documents also raised concerns about how it could be misused by hackers if deployed offensively, which helps explain why it is not currently available to the general public.

Speaking of those recent leaks, it was actually through them that Mythos was first revealed under the name "Capybara", after a draft blog post was exposed in an unsecured cache, something the company later attributed to human error. More recently, Anthropic also exposed Claude Code source files through a packaging mistake, which triggered a messy cleanup effort and accidental GitHub repository takedowns.