Anthropic's New AI Model, Mythos, Enters Limited Cybersecurity Testing

Anthropic has introduced a preview of its latest AI model, Mythos, to a select group of technology and security firms. The model, described in internal documents as the company's most advanced system to date, is being deployed under a program called Project Glasswing. Over forty partners, including Amazon, Apple, Microsoft, and several major cybersecurity vendors, will use Mythos to scan software for vulnerabilities.

The company states that Mythos is a general-purpose model with advanced coding and reasoning abilities, not specifically built for security. Yet in early tests, it reportedly uncovered thousands of previously unknown security flaws, some over a decade old, in proprietary and open-source code. The findings from this private preview will be shared among participants with the goal of improving industry-wide defenses. There are no current plans for a public release.

This launch follows a security lapse last month where details of the model, then internally named 'Capybara,' were leaked from an unsecured data storage system. A separate incident also led to the accidental exposure of hundreds of source code files. Anthropic has acknowledged these as human errors.

The initiative unfolds against a complex regulatory backdrop. Anthropic confirms talks with U.S. federal officials about Mythos, even as the company is engaged in a legal dispute with the Department of Defense over its policies on autonomous surveillance.

Source: TechCrunch