Anthropic Expands Cyber Verification Program After Finding Over 100,000 Vulnerabilities
Anthropic is officially expanding its Cyber Verification Program (CVP) by consolidating Project Glasswing and offering vetted security teams low-guardrail access to advanced Claude models. Between April and October, the initiative helped uncover over 134,000 verified software vulnerabilities, with more than 33,000 rated as critical or high severity. This expansion demonstrates the growing capability of advanced AI models in automated software security and vulnerability research. By partnering with governments and major organizations, Anthropic aims to fortify critical software systems before bad actors can leverage AI for offensive exploits. The updated CVP introduces three access tiers—Defensive, Red Team, and Specialized—with the highest level dedicated to critical infrastructure like power grids and banking systems, co-vetted by the U.S. government. Approved security researchers will gain access to models including Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1 with relaxed safety blocking classifiers.
## BACKGROUND
Advanced LLMs are usually restricted by strict safety classifiers to prevent abuse, such as generating malicious exploit code or facilitating cyberattacks. Anthropic introduced Project Glasswing and the original Cyber Verification Program to grant trusted cybersecurity organizations low-guardrail access, enabling them to discover and patch vulnerabilities in critical open-source and proprietary software.