Anthropic Keeps Powerful New AI Model Under Wraps, Tech Rivals Unite to Beat Hackers to the Fix

Anthropic announced Tuesday that it has developed an AI model capable of finding software vulnerabilities at scale. However, the company is refusing to release it publicly. Instead, the company is granting access to a coalition of rivals—including Apple, Google, Microsoft, and dozens of other tech giants to collectively shore up defenses before malicious actors can exploit the technology.

The model, called Claude Mythos Preview, has already identified thousands of zero-day vulnerabilities across every major operating system and web browser, according to the company. Some of the flaws had gone unpatched for decades, representing a "step change" in AI cyber capabilities, according to Anthropic's internal assessments.

"We see a fundamental shift in the threat landscape," said Dario Amodei, CEO of Anthropic. "Cyberattack timelines are shrinking from months to minutes. We believe the only way to stay ahead is to ensure the good guys have access to these capabilities before the bad guys do."

An Unprecedented Coalition

Dubbed Project Glasswing, the 12 founding partners include Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic has also extended access to more than 40 additional organizations that build or maintain critical software.

The group will use Mythos Preview to find and patch vulnerabilities in shared software infrastructure, conduct black-box testing of binaries, and run coordinated penetration testing. Anthropic has committed up to $100 million in usage credits for the initiative, plus $4 million in direct donations to open-source security organizations, according to people familiar with the matter.

The move follows a March leak that exposed internal Anthropic documents describing Mythos as "far ahead of any other AI model in cyber capabilities" and posing "unprecedented cybersecurity risks." Anthropic confirmed the leak at the time but declined to comment further until Tuesday's formal announcement.

"We are essentially witnessing the Manhattan Project of cybersecurity," said David Gewirtz, a senior editor at ZDNET who serves on the FBI-affiliated InfraGard's Artificial Intelligence Threat and Mitigation Cross-Sector Council. "The question isn't whether AI will redefine cyber warfare—it's whether defenders can coordinate fast enough to prevent chaos."

Long-Hidden Bugs Uncovered

During internal testing, Mythos Preview uncovered a critical flaw in OpenBSD that had gone undetected for 27 years. In another test, it found a 16-year-old vulnerability in FFmpeg, the open-source multimedia framework used by billions of devices. Both flaws, now patched, could have allowed attackers to execute arbitrary code remotely.

The model identified exploitable bugs "better than all but the most skilled humans," according to Anthropic's technical documentation. Security researchers have long warned that AI could dramatically accelerate cyberattacks—Mythos Preview appears to validate those fears.

"It is likely that adversarial actors are already developing similar capabilities," said one senior researcher at a major security firm who asked not to be named due to the sensitivity of the matter. "The window for proactive defense is narrowing fast."

Anthropic says it has no plans for a public release of the model, citing concerns that open access could accelerate AI-powered hacking on a global scale. Instead, the company is betting that unifying competitors around a common defense is the only viable path forward.