Anthropic expands powerful AI model testing program for more security teams

Anthropic has expanded access to its most sophisticated artificial intelligence (AI) models for cybersecurity specialists. This expansion allows previously validated groups to operate with fewer restrictions regarding certain digital security uses.

This change integrates a revised version of the Cyber Verification Program (CVP). The objective of the CVP is to enable specialized entities to use Anthropic's powerful models in vulnerability detection, incident management, and security testing.

This decision was motivated by a previous company initiative called Project Glasswing. This project helped partners discover at least 129 thousand verified software flaws between April and July 2026. Of these discoveries, more than 33 thousand were categorized as critical or high severity.

However, Anthropic emphasizes that these numbers likely represent only a fraction of the total, estimating that the real impact could be at least five times greater, given that the data was collected from only a portion of program participants.

Access Levels and Applications

The first tier, named Defense, is aimed at defensive tasks. Activities covered include malware analysis, vulnerability investigation, and incident response. Candidates for this level can be security teams, essential infrastructure operators, open-source project maintainers, and researchers with a proven track record of identifying and disclosing flaws.

Red Team

The second level, known as Red Team, increases possibilities, allowing for authorized penetration tests and red teaming exercises. To participate at this level, the condition of being an organization is required.

The third level, Specialized, will have the fewest limitations and will be intended for a small group of organizations authorized to test systems considered vital for security, such as power grids, aviation systems, and infrastructure used in interbank transfers.

All members of each category undergo rigorous verification processes. Anthropic informs that it conducts this evaluation in collaboration with the United States government.

The results achieved by Project Glasswing served as justification for the program's expansion. Between April and July, the initiative's partners identified a minimum of 129 thousand confirmed vulnerabilities. Additionally, Anthropic itself located another 5.5 thousand through its open-source code scans conducted between April and October, totaling over 134 thousand verified vulnerabilities. Of these, more than 33 thousand have already been classified as severe or critical.

The company stresses that the actual volume of flaws could be considerably higher, as the Glasswing data depends on information provided by a limited number of partners.

This decision also highlights a central dilemma in using advanced AI models in cybersecurity. The same capabilities that allow a system to locate a vulnerability and assist an expert in fixing it can be used to exploit that flaw.

Anthropic had already addressed similar concerns when launching Claude Mythos in April. At that time, there were apprehensions that AI systems might infiltrate software before their vulnerabilities were patched.

For this reason, the company is not simply removing all protections from its models for any user. Reduced restriction access will only be granted to organizations and professionals who have passed verification processes, under different degrees of control. The intention is to put more advanced capabilities into the hands of those who actively work to find and fix flaws before they are exploited by criminals.