An AI model that can independently discover thousands of software vulnerabilities, some hidden for decades, across major operating systems and web browsers. That’s not the plot of a Netflix thriller. It’s what Anthropic shipped on April 7, 2026, when it unveiled Claude Mythos Preview.
The model’s autonomous cybersecurity capabilities are significant enough that Anthropic immediately restricted access through a new initiative called Project Glasswing, limiting availability to select organizations. No plans for a public commercial release have been disclosed.
What Mythos actually does
Mythos can scan across a variety of major operating systems, web browsers, and other software to identify these hidden flaws. Some of the vulnerabilities it surfaced date back decades, meaning they’ve been quietly sitting in code that billions of people use every day.
That duality is precisely why Anthropic built Project Glasswing as a gatekeeper. The company has prioritized defensive applications, essentially positioning itself as the responsible steward of a technology that could just as easily be weaponized.
