What Project Glasswing and Claude Mythos Preview Are
Project Glasswing is Anthropic’s collaborative security initiative that gives trusted organizations controlled access to Claude Mythos Preview, an advanced vulnerability detection AI that scans codebases for serious flaws and supports secure software development at scale. In practice, this means approved partners can submit large volumes of source code and receive detailed findings on high- and critical-severity bugs, misconfigurations, and exploitable weaknesses. The pilot phase launched with about 50 organizations earlier in the year, using Mythos-class models Anthropic says are more capable for coding tasks than its public Opus family. These early partners have already used Claude Mythos Preview to find more than 10,000 serious vulnerabilities across major operating systems and web browsers, proving that AI security tools can match or exceed expert human reviewers in software vulnerability detection while operating continuously and at enterprise scale.
Project Glasswing Expansion: 150 New Partners, 15 Countries
Anthropic is now widening Project Glasswing by adding approximately 150 new organizations from more than 15 countries to the program. This Project Glasswing expansion follows several weeks of work with the initial cohort, security vendors, open-source maintainers, and government stakeholders. The new partners extend the initiative beyond its first wave of big technology and financial players into sectors such as power, water, healthcare, communications, and hardware. Many are vendors that maintain codebases relied on by governments, companies, nonprofits, and other institutions, which increases the downstream impact of improved security. Anthropic estimates that for most Project Glasswing partners, a major attack on their codebase could affect over 100 million people, with implications for global and national security. Each new participant must meet Anthropic’s security requirements before gaining access, reflecting concern about powerful AI tools being misused by bad actors.
From Finding Bugs to Fixing Them at Scale
The expanded program shows how vulnerability detection AI is changing the balance of effort in cybersecurity. With Claude Mythos Preview, the chronic problem is no longer discovering software flaws but keeping up with verifying, disclosing, and patching them. Some Project Glasswing partners now use the same models not only to scan code but also to write patches, perform pre-release checks, and reduce the chance of introducing new vulnerabilities during development. Anthropic has also released Claude Security, a separate service built on its public models, to scan codebases and suggest fixes, and is providing certain internal tools on request to trusted security teams. Together, these AI security tools point to a near future in which Mythos-class systems can continuously monitor large, shared codebases, triage issues, and propose repairs, while human engineers focus on reviewing changes and coordinating safe releases rather than manually hunting for every bug.
Democratizing AI-Powered Security for Developers and Enterprises
For developers, security teams, and educators, the expansion of Project Glasswing offers a template for how advanced AI can be folded into day-to-day workflows. Anthropic positions the program as part of a broader effort to build AI skills in secure coding, code review, penetration testing, and cyber incident response. By approving organizations of different sizes and sectors, the partnership model brings enterprise-grade software vulnerability detection to groups that might not otherwise access Mythos-class tools. Participants share methods and best practices so that their approaches can be replicated more widely across the industry. As Anthropic notes, AI models are now capable of matching all but the most skilled humans at finding and exploiting vulnerabilities, so controlled initiatives like Glasswing aim to tilt that capability toward defense by giving responsible actors earlier, safer access to Claude Mythos Preview and related AI security tools.
