Anthropic Expands Access to Claude Mythos AI via Project Glasswing

Anthropic is expanding access to its Claude Mythos AI model through Project Glasswing, a program meant to let tech and security firms and governments discover and fix bugs and potential exploits before the powerful model is publicly released.

In a blog post on Tuesday, Anthropic said it is adding roughly 150 organizations to its controlled program as the company prepares for a broader release of the controversial model.

The expansion comes after Anthropic said it expects to bring Mythos-class models to customers "in the coming weeks" once it completes additional safeguards. The news also comes as Anthropic prepares to go public after confidentially filing for an IPO on Monday with the U.S. Securities and Exchange Commission.

“Following several weeks of close collaboration with our Project Glasswing partners, the security industry, open-source software maintainers, and the U.S. government, we’re extending the partnership to approximately 150 new organizations,” Anthropic wrote. “Each one will need to meet our security requirements before they gain access.”

The latest Project Glasswing expansion adds organizations from critical infrastructure sectors, including power, water, healthcare, communications, and hardware. Anthropic said many of the new organizations maintain software used by governments and hundreds of millions of people.

“What each partner has in common is that a successful attack on their codebase could be catastrophic,” Anthropic wrote. “For most partners, we estimate that a major attack could affect more than 100 million people, with important ramifications for both global and national security.”

Claude Mythos emerged publicly in March after draft Anthropic materials leaked online. According to Anthropic, organizations that previously received access have already used Mythos Preview to identify more than 10,000 high or critical-severity software vulnerabilities. Partners are also using the model to write patches, conduct security reviews, and test software before release.

The model's capabilities have raised concerns among researchers and government agencies that increasingly capable AI systems could make cyberattacks easier to execute.

The U.K. AI Security Institute reported that Mythos autonomously completed a 32-step simulated corporate network attack during testing. Mozilla said the model identified 271 vulnerabilities that it fixed in the Firefox web browser, while security startup Calif claimed a preview version helped researchers develop an exploit chain targeting Apple's M5 computer chips.

Anthropic said it is expanding access to Mythos because it believes defenders need increasingly capable AI tools before attackers gain similar capabilities. The company said it expects other developers to release Mythos-class models within six to 12 months, potentially without comparable safeguards, raising the prospect of more frequent and less predictable cyberattacks.

“In the future, frontier model releases will become increasingly high-stakes. Capabilities will continue to improve across all domains, including many that—like cybersecurity—can empower attackers and defenders alike,” Anthropic wrote. “If we’re successful, we hope to enable a permanent advantage for defenders.”

Anthropic did not immediately respond to a request for comment by Decrypt.