Anthropic on Tuesday unveiled "Project Glasswing," an initiative that will let a group of major technology and cybersecurity companies preview an unreleased general-purpose AI model for defensive security applications. The AI startup said the program gives select partners access to its advanced model, called "Claude Mythos Preview," so they can use it in work aimed at improving cybersecurity defenses.
The company named a roster of participating organizations that includes Amazon.com, Microsoft and Apple, alongside cybersecurity firms CrowdStrike and Palo Alto Networks, and technology companies Google and Nvidia. Anthropic said these launch partners will employ Mythos Preview in defensive security efforts and that it will share results and findings with the industry.
Anthropic said Mythos Preview has identified "thousands" of major vulnerabilities in software ranging from operating systems to web browsers. The startup said it will broaden access beyond the core launch partners by extending usage to roughly 40 additional organizations that are responsible for critical software infrastructure.
As part of the program, Anthropic committed up to $100 million in usage credits and pledged $4 million in donations to open-source security groups. The company said its long-term objective is for users to safely deploy Mythos-class models at scale.
The announcement follows media reporting last month that Anthropic had been testing Claude Mythos, a report that described both advanced capabilities and potential security risks associated with the model. That coverage coincided with sharp declines in shares of some cybersecurity firms, including Palo Alto Networks and CrowdStrike, according to Anthropic's account of market reaction.
Industry discussion about AI-driven threats has been prominent at recent security gatherings. Anthropic noted that this year's RSA cybersecurity conference in San Francisco was dominated by conversations about the rise of AI-powered cyberattacks and questions about whether traditional security tools are adequate.
Anthropic also said it has been in ongoing discussions with the U.S. government regarding the model's capabilities. The startup reiterated past disclosures that hackers exploited vulnerabilities in its Claude AI last year to attack approximately 30 global organizations. In addition, the company cited findings from a survey conducted by IBM and Palo Alto Networks that reported 67% of 1,000 executives polled said they had been targeted by AI attacks within the previous year.
Context and next steps - Anthropic plans to collaborate with its partners to apply Mythos Preview to defensive security tasks and to disseminate learnings across the industry. The company is expanding limited access to additional organizations tied to critical infrastructure and has allocated financial resources to support usage and open-source security work.