Anthropic Expands Cyber Verification Program Into 3 Tiers

The program unlocks models like Claude Opus 5.5 and Mythos 5.1 with relaxed malware filters for vetted defenders, while keeping mandatory prompt logging and barring solo researchers from red team access.

Anthropic has overhauled its Cyber Verification Program to give vetted cybersecurity professionals direct access to its frontier models, including Claude Opus 5.5, Sonnet 5.5, and Claude Mythos 5.1. The update, announced in an official company release and reported by SecurityWeek and The Register, rolls the earlier Project Glasswing into a three-tiered authorization framework. Its primary aim is allowing defenders to dissect malicious binaries without triggering standard automated safety refusals.

The Dual-Use Dilemma and Filter Friction

Frontier artificial intelligence models face a constant operational dilemma: dual use. The same reasoning power that untangles complex software vulnerabilities in minutes can help craft working exploits when misapplied. Standard commercial versions of Claude routinely reject requests that touch on malware decompilation, intrusion tools, or reverse engineering.

For security operations centers (SOCs) and defensive engineers, those generic boundaries created daily friction. Legitimate analysts parsing suspect payloads met blanket refusals from the system. The restructured initiative tries to solve that bottleneck by easing automated classifiers exclusively for vetted practitioners and registered organizations.

The Three Tiers of the Cyber Verification Program

The revised framework splits participants across three access tiers based on organizational scope and threat exposure:

• Defense Access: Tailored for defensive tasks such as incident response, vulnerability triage, and malware analysis. This tier is open to vetted individual researchers with a proven public record of responsible disclosure, alongside enterprises, government agencies, universities, and open-source project maintainers.
• Red Team Access: Combines all defensive privileges with permissions for penetration testing and adversary simulation against pre-authorized targets. Anthropic restricts this tier to organizations, barring independent researchers. Real-time safety guardrails stay active to prevent destructive activities, such as deploying ransomware or disrupting physical operations.
• Specialized Access: The tier with the lowest level of automated safety filtering, reserved for vetted entities conducting technical audits on high-impact critical infrastructure, including power grids, telecommunications networks, air traffic control, and interbank transaction systems.

Multi-Cloud Access and Data Logging Requirements

Participants can access the program through the Claude Platform, Google Cloud Vertex AI, and Microsoft Foundry. Availability on Amazon Bedrock depends on the upcoming Enterprise Frontier Safeguards package, scheduled to launch later this fall.

Data privacy remains the main point of contention for technical teams. Anthropic requires mandatory logging of prompts and model interactions to audit for potential policy abuse. A true zero-data-retention setup (keeping telemetry fully isolated on the customer side) will arrive with Enterprise Frontier Safeguards later this fall; for now, zero retention remains restricted to legacy contracts for Fable and Mythos that already carried the waiver.

Portal Application and Timeline

Prospective teams apply through Anthropic’s central Verification Portal, submitting one application per organization. Anthropic projects a review turnaround of up to seven business days, assigning approved applicants to the relevant tier after evaluating submitted credentials.

Sources

• Cyber Verification Program — Anthropic
• CVP Announcement Post — Anthropic on X
• Cyber Verification Program Guidelines — Anthropic Help Center
• Anthropic Introduces 3-Tier Cyber Verification Program for AI Access — SecurityWeek
• Anthropic Reconfigures Its Cool Kids Security Program — The Register