Anthropic opens three-tier AI security access program

Anthropic announced on October 6 an expanded cybersecurity access program that combines its Cyber Verification Program and Project Glasswing, giving verified security professionals and organizations broader access to its high-performance AI cybersecurity capabilities, according to dailysecu.com.

The program is structured in three tiers: defensive, red team, and specialized. Each tier grants access to different Claude models, including Opus 5.5, Sonnet 5.5, and Mythos 5.1. The red team tier covers authorized penetration testing and attack simulation, while actions such as ransomware deployment and physical damage remain blocked.

Individual researchers can apply for the defensive tier based on their vulnerability reporting records. The red team tier is currently open only to organizations.

Anthropic stated that Project Glasswing participants identified at least 129,000 software vulnerabilities between April and July 2026, and that its own open-source review found an additional 5,500 vulnerabilities between April and October 2026. More than 33,000 of the combined findings were rated critical or high severity.

The company noted that vulnerability classification methods vary by organization and that fewer than half of participants disclosed patch numbers, making it difficult to determine how many issues were fixed or posed real risk.

Independent tracking and code security findings

A separate public list managed by VulnCheck researcher Patrick Garrity tracks 300 Anthropic-related CVEs, of which only 2 are marked as exploited in actual attacks. In October 2026, VeraCode's 2026 Generative AI Code Security Report showed a model average security pass rate of 56%, compared with 55% in its previous report; about 44% of tested code generation tasks contained known security vulnerabilities.