Anthropic unveils 'mythos' ai – a cybersecurity shield for tech giants

Silicon Valley is bracing for a new era of cyber defense, thanks to Anthropic’s preemptive move. The AI firm is quietly arming major tech players—Amazon, Apple, Microsoft, and Cisco—with access to its unreleased 'Mythos' model, designed to proactively hunt for vulnerabilities before malicious actors exploit them. The initiative, dubbed Project Glasswing, signals a growing anxiety within the industry about the escalating cybersecurity risks posed by increasingly sophisticated AI.

A race against emerging threats

The revelation follows a recent leak of a blog post detailing Mythos, and arrives as OpenAI similarly acknowledged the escalating cyber capabilities of its own models, launching a pilot program to get its tools into the hands of security professionals first. Anthropic’s proactive stance suggests a deep understanding of the impending challenges. Newton Cheng, Frontier Red Team's cybersecurity lead, emphasized that this isn't solely Anthropic's concern. “It’s an industry-wide problem that both private companies and governments must address,” he stated, framing Glasswing as a means to give defenders a crucial head start.

The model itself, Anthropic claims, has already demonstrated remarkable aptitude. It flagged a security flaw dating back 27 years in critical internet software, a vulnerability that slipped past five million automated tests. Even more strikingly, Mythos identified a 16-year-old coding error in a popular video game—a flaw that automated testing tools had completely missed. Dianne Penn, Anthropic’s head of product management for research, underscored the stringent access controls in place within Project Glasswing, though details remain understandably scarce for security reasons.

While Mythos isn't specifically built for cybersecurity, its general-purpose nature has yielded unexpectedly potent results. Anthropic has reportedly engaged with U.S. government agencies regarding Mythos’ security capabilities, collaborating currently with CISA and NIST. The fact that a model not explicitly designed for security can uncover such deeply embedded vulnerabilities underscores the potential—and the peril—of advanced AI.

The leak, though embarrassing, may have inadvertently accelerated the discussion around AI security. It's a clear indication that the race to secure the future of Technology is already underway, and Anthropic’s Project Glasswing represents a significant, if secretive, gambit in that contest.