As frontier AI reaches unprecedented hacking capabilities, tech giants unite to turn the ultimate vulnerability scanner into the ultimate defensive shield.
- A Paradigm Shift in Capability: Anthropic’s unreleased Claude Mythos Preview model has demonstrated an unparalleled ability to find and exploit high-severity vulnerabilities, surpassing all but the most skilled human experts and prompting a decision to withhold it from general public release.
- The Birth of Project Glasswing: To preemptively secure critical software before these advanced capabilities proliferate, an unprecedented coalition of industry leaders—including Apple, Google, Microsoft, and NVIDIA—has formed to utilize this AI strictly for defensive cybersecurity.
- A Massive Defensive Investment: Backed by up to $100 million in usage credits and $4 million in direct donations for open-source security, this initiative aims to keep global cyber defenders one step ahead of an accelerating technological curve.
We are standing at the precipice of a new era in cybersecurity, driven by the rapid and staggering advancement of artificial intelligence. Recent developments have revealed a stark reality: AI models have now reached a level of coding capability where they can surpass all but the most highly skilled human experts at finding and exploiting software vulnerabilities. This reality was brought into sharp focus by Claude Mythos Preview, a general-purpose, frontier AI model developed by Anthropic. Boasting a striking leap in benchmark scores compared to its predecessor, Claude Opus 4.6, Mythos Preview has already uncovered thousands of high-severity vulnerabilities embedded across every major operating system and web browser.
Given the sheer velocity of AI progress, it will not be long before models with this level of capability become widespread. The proliferation of such power, especially if it falls into the hands of bad actors, carries severe potential fallout for global economies, public safety, and national security. Recognizing this immense risk, Anthropic conducted extensive safety evaluations, detailed in a comprehensive System Card. These assessments covered their Responsible Scaling Policy, Frontier Compliance Framework, cybersecurity skills, alignment, model welfare, and user experiences. The results led to a critical decision: Claude Mythos Preview will not be made generally available. Instead, the goal is to figure out how to deploy Mythos-class models safely at scale, a process that requires reliable safeguards to block their most dangerous outputs.
To meet this challenge head-on, Anthropic and a coalition of industry titans have launched Project Glasswing—an urgent, proactive attempt to put these cutting-edge capabilities to work exclusively for defensive purposes. This initiative brings together an unprecedented roster of launch partners, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. By providing these launch partners with access to Mythos Preview, the project aims to secure the world’s most critical software infrastructure from the inside out.
The scope of Project Glasswing extends far beyond its massive corporate partners. Access to the model has been expanded to a group of over 40 additional organizations that build and maintain critical software, enabling them to scan and secure both first-party and open-source systems. To fuel this massive defensive undertaking, Anthropic is committing up to $100 million in Mythos Preview usage credits, alongside $4 million in direct donations to open-source security organizations. Furthermore, Anthropic plans to begin testing necessary safety safeguards with an upcoming Claude Opus model, using the findings from the Mythos Preview System Card to inform the release and security protocols of all future Claude iterations.
Project Glasswing is just a starting point. Defending the world’s cyber infrastructure is a monumental task that will likely take years, yet frontier AI capabilities are projected to advance substantially in just the next few months. No single organization can solve these existential cybersecurity problems alone. It requires a unified effort where frontier AI developers, software companies, security researchers, open-source maintainers, and governments worldwide all play an essential role. Anthropic has committed to reporting back on what is learned throughout this process so that the entire industry can benefit. For cyber defenders to come out ahead in the AI era, the time to act is now.
