In a strategic move to proactively address the evolving landscape of artificial intelligence security, AI safety leader Anthropic has launched a preview of its groundbreaking, unreleased frontier model, codenamed Claude Mythos. This initiative, dubbed Project Glasswing, circumvents a traditional public release, instead offering early access to a curated group of prominent technology and cybersecurity partners. The decision stems from a heightened awareness of potential risks associated with deploying highly capable AI models, a concern brought into sharper focus by a recent security lapse within Anthropic itself.
The genesis of this cautious approach can be traced back to late March, when a misconfiguration within Anthropic’s content management system inadvertently exposed information about Claude Mythos. This accidental leak revealed that the company was developing a new tier of AI models intended to surpass the capabilities of its current flagship, Opus. The unannounced details also highlighted Anthropic’s intention to adopt a more deliberate launch strategy to mitigate "potential near-term risks in the realm of cybersecurity." This foresight is now being actively implemented through Project Glasswing.
Project Glasswing: A Collaborative Approach to AI Defense
Claude Mythos Preview, described by Anthropic as a "general-purpose, unreleased frontier model," is not yet available to the general public, nor will it be in its current iteration. Instead, Anthropic is extending access to a select consortium of partners through Project Glasswing. This initiative aims to leverage the advanced capabilities of Mythos for defensive security purposes, empowering organizations to bolster their own systems and the open-source tools they rely upon.
The inaugural partners for Project Glasswing represent a formidable coalition in the technology and security sectors. These include industry giants such as Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. These organizations, along with approximately 40 other entities, will gain access to the preview version of Claude Mythos. Their primary objective will be to utilize the model for scanning and securing their internal infrastructure and the vast ecosystem of open-source software.
While Claude Mythos was not explicitly designed as a cybersecurity tool, Anthropic has observed its exceptional performance in agentic coding and reasoning tasks. This proficiency has been validated by its performance on the CyberGym benchmark, a platform that evaluates AI agents on their ability to conduct vulnerability analysis. In these assessments, Claude Mythos achieved an impressive score of 83.1%. This significantly outperforms its predecessor, Opus 4.6, which previously held the top position on this benchmark with a score of 66.6%. The substantial leap in performance underscores the potential of Mythos to revolutionize security analysis.
A Proactive Stance on AI Safety and Security
Anthropic, founded with a strong emphasis on AI safety as an alternative to other leading AI research labs, appears committed to a philosophy of responsible AI deployment. The company’s decision to release a model with such advanced capabilities through a controlled preview reflects a belief that defenders should be equipped to counter potential threats before attackers gain access to similar technologies. This "security-first" approach aims to prevent a scenario where sophisticated AI tools are weaponized by malicious actors.
While some might perceive this emphasis on potential dangers as a marketing strategy to enhance the perceived value of these advanced models, the underlying security risks are undeniable. Elia Zaitsev, Chief Technology Officer at CrowdStrike, articulated the urgency of the situation in a recent statement. He highlighted the dramatic compression of the vulnerability exploitation timeline, noting, "The window between a vulnerability being discovered and being exploited by an adversary has collapsed – what once took months now happens in minutes with AI." Zaitsev further emphasized the dual nature of this development: "Claude Mythos Preview demonstrates what is now possible for defenders at scale, and adversaries will inevitably look to exploit the same capabilities. That is not a reason to slow down; it’s a reason to move together, faster. If you want to deploy AI, you need security."
Real-World Impact: Uncovering Critical Vulnerabilities
The efficacy of Claude Mythos Preview in identifying security weaknesses has already been demonstrated. Anthropic has reported using the model to uncover "thousands of zero-day vulnerabilities, many of them critical." These discoveries span a range of software, including a 27-year-old, previously unknown, and unpatched bug in OpenBSD. In another significant finding, the model successfully chained multiple vulnerabilities within the Linux kernel to achieve superuser access, showcasing its sophisticated exploit chain discovery capabilities.
The implications of these findings are particularly profound for the open-source community. Jim Zemlin, Executive Director of the Linux Foundation, underscored the historical disparity in security resources. He stated, "Open source maintainers – whose software underpins much of the world’s critical infrastructure – have historically been left to figure out security on their own." Zemlin elaborated on the transformative potential of Project Glasswing: "By giving the maintainers of these critical open source codebases access to a new generation of AI models that can proactively identify and fix vulnerabilities at scale, Project Glasswing offers a credible path to changing that equation. This is how AI-augmented security can become a trusted sidekick for every maintainer, not just those who can afford expensive security teams." This initiative aims to democratize advanced security capabilities, making them accessible to projects that may lack the financial resources for dedicated security teams.
Future Outlook and Financial Commitments
While Anthropic eventually plans to make Mythos-class models available to its broader user base, the current focus remains exclusively on Project Glasswing participants. To support this collaborative effort, Anthropic is providing $100 million in usage credits to the participating companies. Furthermore, the company is allocating $4 million in direct donations to open-source security organizations, demonstrating a commitment to strengthening the foundational elements of the digital infrastructure.
This strategic rollout underscores Anthropic’s commitment to responsible AI development and deployment. By prioritizing security and collaborating with key industry players, Project Glasswing aims to not only advance the capabilities of AI but also to ensure its safe and beneficial integration into critical systems. The lessons learned and vulnerabilities identified through this preview are expected to shape the future of AI-powered cybersecurity and the broader AI landscape. The initiative represents a significant step towards a future where AI acts as a powerful ally in the ongoing battle for digital security, empowering defenders to stay ahead of evolving threats. The success of Project Glasswing could set a precedent for how future advanced AI models are introduced, emphasizing collaboration and proactive risk mitigation.