Anthropic Unveils 3-Tier Cyber Verification Program for AI Access
Anthropic has integrated its Cyber Verification Program and Project Glasswing into a unified framework offering three distinct tiers of access to its most capable AI models.
Radwan AbdulhadiEditor and technology writer
Unified Program Structure
Anthropic announced a revamped Cyber Verification Program (CVP), combining CVP and Project Glasswing into a single unified offering with three distinct tiers of access to its most powerful artificial intelligence models. Generally available models like Claude Opus 5.5, Sonnet 5.5, and Fable 5.1 normally ship with strict cyber safeguards that block most cyber security tasks because the same underlying capabilities that help defenders identify flaws can also be leveraged by attackers to exploit them.
Previously, CVP and Glasswing operated as separate initiatives, where Glasswing provided organizations securing critical software with access to Claude Mythos, while the original CVP relaxed safeguards on Opus and Sonnet models for pre-approved teams. Under the new structure, all three tiers incorporate Opus 5.5, Sonnet 5.5, Mythos 5.1, and future models, with each tier governed by specific verification requirements and security controls.
Defense Access Tier
The Defense Access tier is engineered for security operations center (SOC) and incident response work, malware reverse engineering, vulnerability analysis, and validation. Eligible applicants for this tier include internal security teams defending their own networks, critical infrastructure operators, small security firms, open source maintainers, and individual researchers who possess a documented history of reported vulnerabilities.
Anthropic stated that it aims to process and respond to applications for the Defense Access tier within a few days. Organizations accepted into the new program are required to accept data retention terms so that Anthropic can actively monitor for system misuse, though certain zero data retention options remain available under specific configurations.
Red Team Access and Specialized Access
The Red Team Access tier introduces authorized penetration testing and red teaming capabilities, strictly limited to systems that the participating organization has explicit permission to test. However, actions that could trigger mass disruption or physical harm, such as deploying ransomware, remain actively blocked in real time by system safeguards.
Anthropic noted that this intermediate tier is currently restricted to organizations only, meaning individual researchers are not eligible. Review processes for Red Team Access are expected to take a few weeks, with qualifying applicants granted Defense Access in the interim. Meanwhile, the Specialized Access tier features the fewest cyber blocks and is strictly reserved for a select group of organizations authorized to test safety-critical systems like power grids, telecom networks, flight systems, and interbank transfer infrastructure.
Vulnerability Findings and Platform Availability
Data shared alongside the program update highlighted extensive vulnerability discoveries. Glasswing partners uncovered at least 129,000 verified vulnerabilities between April and July, while Anthropic's internal open source scanning identified an additional 5,500 flaws between April and October. Over 33,000 of these total findings were rated as critical or high severity, with Anthropic estimating the true impact is likely at least five times higher given that the figures derive from only a subset of participants.
Regarding availability, the updated CVP is accessible across the Claude Platform, Google Cloud’s Vertex AI, and Microsoft Foundry. On Amazon Bedrock, availability is restricted strictly to customers who qualify for Enterprise Frontier Safeguards. Existing CVP members maintain their current settings for legacy models while being automatically evaluated for access to the newer offerings.
Sources
- SecurityWeekAnthropic Introduces 3-Tier Cyber Verification Program for AI Access
Continue chronologically
Related entity coverage
Related stories
Anthropic Asks Claude Users to Share Voice Data
Anthropic is asking Claude users to voluntarily share their voice conversations and audio recordings to help train and improve its AI models.
Amazon Bedrock Adds Anthropic In-Region Inference in Seoul and Singapore
Amazon Bedrock now supports Anthropic models with in-region inference in Seoul and Singapore for strict local data processing requirements.
Claude Code and Amazon Bedrock in AWS GovCloud
Discover how Anthropic Claude Opus 5.5 and Claude Sonnet 5.5 on Amazon Bedrock enable compliance-aligned, AI-assisted development for regulated workloads.
Add Secure Web Search to Claude Desktop with Amazon Bedrock AgentCore
Learn how to connect Claude Desktop to secure Web Search using Amazon Bedrock AgentCore Gateway and JWT-based authentication.
