Anthropic is expanding its Cyber Verification Program to give more vetted cybersecurity teams access to its most powerful Claude models with fewer safety filters, after partners used Project Glasswing to uncover at least 129,000 verified software vulnerabilities between April and July 2026[1][3][13].
The revamped program folds Project Glasswing and Anthropic’s earlier Cyber Verification initiative into a single, three-tier framework that opens Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1 to qualifying organizations under different levels of scrutiny[6][9][13]. Anthropic and its partners report that more than 33,000 of the 129,000 verified flaws have been rated critical or high severity, underscoring the potential of AI-assisted analysis against systemically important software[3][8][14]. In addition, Anthropic’s own scanning of open-source code has identified roughly 5,500 further vulnerabilities through October 2026, many in projects that underpin modern infrastructure and cloud services[1][14].
Under the new structure, a Defense tier supports incident response and malware analysis, a Red Team tier layers on authorized penetration testing, and a Specialized tier grants the most permissive access for safety-critical domains such as power grids, aviation and financial transfer systems[6][13]. Membership in the Specialized tier is reserved for a small set of organizations vetted jointly by Anthropic and the U.S. government, with existing Project Glasswing participants slated to migrate into this group[6][13][14]. All tiers provide reduced content filters and blocking classifiers compared with standard Claude deployments, allowing experienced security teams to push the models toward realistic exploit scenarios while still operating within defined guardrails[1][6][13].
Project Glasswing itself launched in April 2026 as a collaboration with major technology firms and financial institutions, including cloud providers, chipmakers and banks, focused on securing what Anthropic describes as the world’s most critical software[2][5]. In an initial update earlier this year, Anthropic said its roughly 50 Glasswing partners had already found more than 10,000 high- or critical-severity issues across operating systems, browsers and other foundational software components, out of more than 23,000 vulnerabilities in total[5][7][11]. By June, the company extended Glasswing to about 150 additional organizations, highlighting sustained demand for AI tooling that can interrogate widely deployed platforms at scale for previously unknown flaws[7][11].
Outside observers note that the discovery of six-figure vulnerability counts in just a few months illustrates both the depth of latent bugs in production systems and the disruptive potential of large-scale automated code review[8][10][15]. At the same time, expanding access to models optimized for offensive security testing raises familiar dual-use questions, including how to prevent misuse while ensuring that defenders in smaller organizations and critical infrastructure can benefit from advanced AI capabilities.
Anthropic says qualifying entities include security teams at companies, nonprofits, universities and government agencies defending the systems they operate, as well as operators of critical infrastructure such as hospitals and utilities, smaller security firms, open-source maintainers and individual researchers with a track record of responsible vulnerability disclosure[1][6][9]. The company emphasizes that participants are expected to focus on assets they own or are explicitly authorized to test, and to route findings through established disclosure channels to vendors and coordinating bodies, with ongoing oversight from Anthropic and government partners[1][6][12].
References
- Expanding the Cyber Verification Program
- Project Glasswing: Securing critical software for the AI era
- Anthropic Expands Claude Access for Vetted Cyber Teams …
- Project Glasswing: An initial update
- Anthropic Expands Cyber Verification Program to Three …
- Expanding Project Glasswing Anthropic
- Anthropic reconfigures its cool kids security program
- Anthropic opens its most powerful AI models to more …
- Anthropic Expands Project Glasswing, Introduces Three-Tier Access
- Expanding Project Glasswing
- Anthropic expands AI model access for government and …
- Anthropic opens its most powerful AI models to more …
- Anthropic folds Project Glasswing into an expanded three-tier Cyber Verification Program – SiliconANGLE
- Anthropic expands cyber AI access program to more security firms
