What Anthropic Project Glasswing Is and Why This Expansion Matters
Anthropic Project Glasswing is a controlled-access initiative that gives selected organizations the Mythos AI model to strengthen software security, turning advanced AI into a practical tool for finding and fixing critical vulnerabilities in large codebases across key industries. Anthropic is now expanding Project Glasswing to around 150 additional organizations in more than 15 countries, following an initial cohort of about 50 partners that started in April with Claude Mythos Preview. These partners have already identified over 10,000 high- or critical-severity flaws, showing how AI cybersecurity tools can expose systemic weaknesses at scale. The new wave includes banks, regulators, critical infrastructure providers, and enterprises that maintain software relied on by governments, companies, and nonprofits. Each participant must meet Anthropic’s security requirements before gaining access, reinforcing that enterprise AI security at this level is not a general-purpose API, but a selective program with high stakes.
From Vulnerability Discovery to Enterprise AI Security Strategy
Project Glasswing’s focus on the Mythos AI model highlights how frontier systems are moving beyond generic coding help into specialized enterprise AI security. Mythos is being used for code review, vulnerability detection, penetration testing, and secure development, pointing to a future where AI agents routinely scan and reason about massive software estates. Anthropic estimates that for most partners, a major attack on their codebase could affect more than 100 million people, underlining why access is limited to vetted organizations. The expansion also extends to sectors that were underrepresented before, including power, water, healthcare, communications, and hardware, where software failures can quickly cascade into physical and national security incidents. Anthropic describes this as a step toward making “all software more secure” and helping the industry adjust as AI changes core cybersecurity assumptions, pushing enterprises to treat AI security expansion as a board-level concern rather than a narrow tooling choice.
AI Cybersecurity Arms Race: Mythos vs Rival Frontier Models
Anthropic’s tighter control over Mythos contrasts with rivals adopting broader distribution. The company restricts access through programs like Project Glasswing and its Cyber Verification Program, arguing that advanced cyber capabilities need stronger safeguards before wide release. At the same time, other developers are racing to supply competing AI cybersecurity tools. Microsoft has introduced its MAI family of in-house models, aiming to compete more directly in enterprise AI infrastructure after reshaping its relationship with OpenAI. OpenAI itself is pushing ahead with GPT-5.5 Cyber, which has already been granted to several major banks. According to the UK AI Security Institute, Claude Mythos Preview solved a complex 32-step simulated enterprise attack chain known as TLO, and later tests found an OpenAI 5.5 model achieving a 71.4% pass rate on expert-level cyber tasks, compared with 68.6% for Mythos Preview, showing capability is spreading across multiple vendors.
Implications for Banks, Regulators, and Critical Infrastructure Operators
The Glasswing expansion is especially significant for banks, regulators, and operators of critical infrastructure, which sit at the center of the emerging AI security expansion. Some financial institutions have expressed frustration at not having Mythos access, while others are moving ahead with alternative frontier models. Japanese banks are among the new Mythos users, after financial authorities set up a task force to understand AI-driven cyber risks. In parallel, major banks elsewhere are pursuing GPT-5.5 Cyber and other tools. For critical sectors like power, water, and healthcare, Glasswing offers a way to integrate AI cybersecurity tools under close oversight: Mythos can discover large volumes of severe issues, while humans remain responsible for triage, disclosure, and patching. As Anthropic notes, the bottleneck is shifting from finding flaws to fixing and deploying patches, suggesting enterprises will need new workflows, governance, and staffing models to keep pace.
How Anthropic’s Education Push Shapes Enterprise Adoption
Project Glasswing is not only a technical deployment; it also functions as a structured education program for AI skills and cybersecurity practice. Anthropic’s partners include developers, software teams, cybersecurity researchers, and education providers teaching AI and cyber skills, who are collectively testing how frontier models can be folded into existing security operations. Participants share methods, coordinate with third parties to triage findings, and validate how AI-generated vulnerability reports translate into patches. This collaborative model positions Anthropic as a central voice in enterprise AI security, helping define best practices as organizations experiment with Mythos for secure development and code review. For enterprises, the message is that adoption now means more than plugging in a new scanner: it requires training developers to work with AI, revising threat models, and aligning leadership on when and how to trust AI-assisted security judgments in production environments.
