Anthropic has expanded its Cyber Verification Program (CVP), allowing approved cybersecurity professionals and organisations to test advanced artificial intelligence models with reduced safeguards for authorised security research. The company said the expansion follows findings from Project Glasswing, an initiative focused on using AI capabilities to identify software vulnerabilities, which reportedly uncovered at least 129,000 verified security flaws between April and July 2026. Anthropic also reported discovering an additional 5,500 verified software vulnerabilities between April and October 2026 through open source scanning efforts. The company stated that more than 33,000 of the identified vulnerabilities have been classified as critical or high severity, while noting that the actual number may be higher because the data was collected from only a portion of Project Glasswing partners.
The updated Cyber Verification Program introduces three access tiers designed to support different cybersecurity use cases. The Defense Access tier is intended for defensive activities such as incident response, malware analysis, vulnerability assessment and security validation. The Red Team Access tier expands capabilities to include authorised penetration testing and red team exercises alongside defensive security work. A third tier, called Specialized Access, provides the highest level of access with fewer safeguards and is reserved for a limited number of verified organisations that are approved to test AI safety systems. Through these access levels, organisations can use Anthropic models including Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1, along with future models released under the programme.
Anthropic said evaluations conducted through CyScenarioBench showed differences in how security safeguards performed across different access levels. According to the company, safeguards blocked 46 out of 50 evaluated tasks when Claude Opus 5.5 was used under the Defense Access tier. Under the Red Team Access tier, the same model completed 34 out of 50 tasks without blocking any, matching the completion rate observed when safeguards were not applied. Without Cyber Verification Program access, Anthropic said all evaluated tasks were blocked at the first prompt. The company explained that the programme is designed around the dual use nature of advanced AI capabilities, allowing cybersecurity professionals to use similar technologies for system protection while reducing risks associated with misuse. Anthropic said expanding access to verified defenders could strengthen security efforts by helping organisations identify and address weaknesses before they are exploited.
Separately, research from VulnCheck examined vulnerabilities associated with Anthropic and Project Glasswing discoveries, finding that only a small portion of the identified flaws have been exploited in real world attacks. According to researcher Patrick Garrity, two out of 300 analysed vulnerabilities, representing 0.67 percent, had evidence of active exploitation. These included CVE 2026 26980, an SQL injection vulnerability affecting Ghost CMS, and CVE 2026 61500, a session forgery vulnerability impacting Rejetto HTTP File Server. The findings indicate that while AI assisted security research may help accelerate vulnerability discovery, the presence of a vulnerability does not always mean it can be easily exploited or result in significant impact. Security researchers have also highlighted that AI generated code and automated fixes require human review, as some AI assisted development processes may introduce additional security concerns. Research from Veracode has shown that AI generated code can still contain security weaknesses, reinforcing the need for organisations to combine AI capabilities with established software security practices, testing processes and expert oversight.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
