Anthropic expanded access for cybersecurity professionals to its more advanced artificial intelligence (AI) models, allowing previously verified teams to work with less protections against certain uses of digital security.
The change is part of a revamped version of the Cyber Verification Program (CVP), created to allow specialized organizations to use powerful Anthropic models for identifying vulnerabilities, incident response, and security testing.
The decision comes after a previous initiative by the company, called Project Glasswing, helped partners find at least 129 thousand verified software vulnerabilities between April and July 2026. More than 33 thousand of these flaws have been classified so far as critical or high severity.
Anthropic states, however, that the numbers likely represent only a portion of the total. The company estimates the real impact could be at least five times greater, as the data was obtained from a survey with only part of the program's partners.
Fewer barriers for those testing systems
- The new program combines two initiatives that Anthropic had been conducting separately for the last six months;
- The Project Glasswing allowed organizations responsible for the security of critical software to use the Claude Mythos, the company's family of models focused on advanced cybersecurity capabilities;
- The original CVP provided previously verified security teams access to versions with fewer protections of the Claude Opus and Sonnet models;
- Now, the two programs have been combined into a single structure;
- The new CVP will have three levels of access, each with its own verification requirements and security controls;
- All of them include access to Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1, in addition to future Anthropic models.
Three levels of access
The first level, called Defense, is focused on defensive work.
Among the planned activities are incident response, malware analysis, and vulnerability investigation. Security teams, critical infrastructure operators, open-source project maintainers, and researchers with a history of identifying and disclosing vulnerabilities can apply.
The second, Red Team, expands possibilities to include authorized penetration tests and red teaming exercises. In this case, only organizations can request participation.
The third level, Specialized, will have the fewest restrictions. He will be reserved for a small group of authorized organizations to test systems considered critical for safety, such as electric grids, aviation systems, and infrastructure used for interbank transfers.
New program combines two initiatives that Anthropic has been conducting separately for the last six months – Image: Mehaniq/Shutterstock
Members of each category will be subject to verification processes. According to Anthropic, the company conducts this evaluation in conjunction with the U.S. government.
Read more:
- Anthropic: IPO prospectus shows the company is dependent on rivals
- U.S. investigates OpenAI and Anthropic after incidents with AI agents
- How much does the CEO of one of the most important AI companies of the moment earn
AI that finds vulnerabilities
The results of Project Glasswing helped justify the expansion of the program.
Between April and July, the initiative partners identified at least 129,000 verified vulnerabilities. Anthropic itself found another 5,500 through its open-source code scanning efforts conducted between April and October. Together, the initiatives found over 134,000 verified vulnerabilities. More than 33 thousand have already been classified as critical or high severity.
The company highlights that the number could be significantly higher, since the Glasswing data is based on information provided by a limited portion of the partners.
The risk of releasing a more powerful AI
The decision also exposes one of the main dilemmas involving advanced AI models applied to cybersecurity.
The same capabilities that allow a system to find a vulnerability and help an expert fix it could be used to exploit a flaw.
Anthropic had already faced similar concerns when presenting the Claude Mythos in April. At that time, the launch raised fears that AI systems could manage to invade software before its vulnerabilities were fixed.
Therefore, the company is not simply removing protections from its models for any user. Access with fewer restrictions will be directed to organizations and professionals subjected to verification processes, with different levels of control.
The proposal is to place more advanced capabilities into the hands of those who work precisely to find and fix flaws before they are exploited by criminals.
Rodrigo Mozelli is a journalist graduated from the Universidade Metodista de São Paulo (UMESP) and is currently a writer at Olhar Digital.
