Anthropic Overhauls Cyber Verification Program with Three Tiered Access Framework

Restructuring the Cyber Verification Framework

Anthropic unveiled a major expansion of its Cyber Verification Program (CVP) in mid-October 2026. The overhaul alters how authorized cybersecurity practitioners use frontier reasoning models, reducing prompt refusals during legitimate security research and incident remediation.

The update integrates Project Glasswing, Anthropic's previously invite-only research initiative for critical software hardening. Between April and October 2026, Project Glasswing partners and associated vulnerability scanners identified more than 134,000 verified software vulnerabilities, including over 33,000 high and critical severity flaws.

Through this reorganization, Anthropic established a three-tier access hierarchy aligned with specific technical operational profiles.

Three Operational Access Tiers

The access model separates defensive telemetry analysis from active penetration testing:

• Defense Access: Tailored for security operations centers, incident responders, malware reverse engineers, and vulnerability validation teams. Eligible applicants include enterprise defenders, critical infrastructure operators, open source maintainers, and vetted individual researchers.

• Red Team Access: Authorizes penetration testing restricted strictly to infrastructure owned by the participating entity. Real-time safety filters remain active to prevent mass-disruption workflows such as automated ransomware compilation. This tier is restricted to verified organizations.

• Specialized Access: The highest clearance tier, absorbing previous Project Glasswing participants to provide deep model collaboration for national infrastructure defense.

All three tiers grant access to Anthropic's technical models, including Claude Opus 5.5, Claude Sonnet 5.5, and the specialized Claude Mythos 5.1 series.

Balancing Relaxed Refusal Classifiers with Strict Credentialing

A core friction point for security teams using general-purpose models has been excessive false-positive refusals, where defensive requests are incorrectly flagged as malicious exploitation attempts.

To balance reduced blocking classifiers with risk containment, Anthropic implemented strict identity requirements for participating teams:

• Phishing-resistant authentication: Participants must use hardware-based FIDO2 security keys or WebAuthn passkeys to authenticate to program endpoints.

• Deprecation of static API keys: Long-lived static API secrets are replaced with short-lived, platform-brokered tokens.

• Immutable audit trails: All tool calls and prompt transactions are recorded in cryptographically verifiable audit logs to monitor authorized scope compliance.

The CVP rollout coincides with the launch of Anthropic's automated OSS Scanner and the Critical Infrastructure Defense Program, shifting frontier AI deployment toward structured defensive assurance.