Expanding the Cyber Verification Program

Anthropic launched an expanded Cyber Verification Program that establishes three access tiers providing verified security professionals with advanced cyber capabilities and reduced classifier blocks across frontier models. According to Anthropic News's report, the program unifies Project Glasswing and previous verification initiatives, unlocking Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1. While standard models like Claude Opus 5.5, Claude Fable 5.1, and Claude Sonnet 5.5 enforce strict safeguards blocking cyber tasks to prevent misuse, the expanded program tunes safeguards to verified organizational roles.

This report covers the operational mechanics, qualification criteria, benchmark results, and data governance standards governing the expanded offering. It is intended for enterprise defenders, critical infrastructure operators, academic vulnerability researchers, and authorized penetration testers evaluating tier eligibility, turnaround timelines, and zero-data-retention compliance pathways.

Expanding the Cyber Verification Program: what actually changed

The structural modification merges two separate initiatives from the previous six months into a unified program with three distinct tiers. Previously, Anthropic operated Project Glasswing to give organizations protecting critical software access to Claude Mythos, alongside an early verification program providing vetted teams with reduced safeguards on Claude Opus and Claude Sonnet. Under the unified structure, security professionals apply directly for Defense Access, Red Team Access, or Specialized Access based on their technical remit. All three tiers include access to Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1, and future foundation models.

This overhaul replaces ad hoc onboarding with standardized verification controls, differentiated classifier calibrations, and clear eligibility parameters. Generally available models like Claude Opus 5.5, Claude Sonnet 5.5, and Claude Fable 5.1 retain conservative safeguards that block cyber operations to curb malicious actors. Standard models continue handling everyday defensive workflows such as code review, patch application, alert triage, and discovering vulnerabilities in owned source code. For deeper analysis, the updated program reduces automated classifier blocks for vetted teams, while existing members retain their settings and receive automatic evaluations for the new models without submitting redundant applications.

Expanding the Cyber Verification Program: how it works

Security teams access model capabilities through a tiered structure where each level reflects specific operational permissions and vetting depth. Defense Access covers core defensive functions including incident response, security operations center analysis, reverse-engineering malware, and validating vulnerabilities, with application reviews concluding within a few days. Red Team Access adds authorized penetration testing and adversarial simulations against explicitly authorized IT systems in critical industries. Because Red Team Access entails heightened eligibility reviews, applications take several weeks, during which qualifying organizations are enrolled in Defense Access while their red-team applications are processed.

Specialized Access maintains the fewest cyber blocks, reserved for vetted organizations authorized to test systems whose disruption could impact human lives or markets, such as flight operations, power grids, telecom networks, interbank transfers, and government administrative networks. Anthropic reviews Specialized Access applicants in depth alongside the United States government, with existing Project Glasswing members transitioning automatically. Across all tiers, real-time blocks persist against ransomware deployment, damaging physical equipment, or testing high-risk physical safety systems. Platform data retention is required for misuse monitoring, though organizations holding zero-data-retention access for Claude Fable 5.1 or Claude Mythos 5.1 can retain those settings.

Expanding the Cyber Verification Program: why it matters now

Defenders require advanced foundation models to secure complex environments, as evidenced by vulnerability discovery rates under Project Glasswing. Participating partners using Claude Mythos uncovered at least 129,000 verified software vulnerabilities between April and July 2026. Internal open-source scanning efforts surfaced an additional 5,500 verified vulnerabilities between April and October 2026, with more than 33,000 total issues rated as critical or high severity. Partners including Booz Allen and Comcast reported that deploying Claude Mythos accelerated vulnerability finding by months or years. Anthropic treats these figures as an undercount derived from 33 partner reports, expecting true ecosystem impact to be at least five times higher.

Evaluation results on CyScenarioBench demonstrate why tailored classifier levels are essential for practical operations. Testing Claude Opus 5.5 across 10 multi-stage scenarios over 50 trials per tier showed that standard models without program access were blocked on the initial prompt in every instance. In Defense Access, safety filters blocked 46 out of 50 trials, allowing only four tasks to complete. In Red Team Access, Claude Opus 5.5 experienced zero automated blocks and completed 34 of 50 challenges, achieving a 68 percent success rate equivalent to the model's 67.6 percent completion benchmark when running without safeguards under Specialized Access.

Expanding the Cyber Verification Program: who is affected

The program establishes clear eligibility criteria across enterprises, public infrastructure, and individual researchers. Defense Access admits corporate security teams defending internal systems, non-profits, universities, and government bodies. It also includes operators of critical infrastructure such as regional hospitals and municipal utilities, alongside small security firms, open-source maintainers, and individual researchers who have a verifiable record of reporting vulnerabilities. This broad eligibility ensures that frontline defenders across diverse organizational sizes can utilize Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1 for malware analysis, alert triage, and defensive validation.

Participation rules tighten for offensive testing, barring individual researchers from Red Team Access and Specialized Access. Red Team Access admits only organizations, including in-house corporate red teams, government red teams, and authorized penetration testing firms evaluating systems they have explicit permission to test. Specialized Access is restricted to a small group of verified organizations testing life-safety and market-critical systems such as telecom networks, power grids, flight operations, and interbank transfer infrastructure. These organizations undergo extensive collaborative reviews with the United States government before gaining approval to deploy models under minimal classifier restrictions.

Expanding the Cyber Verification Program: what to watch

Security and compliance teams should monitor the rollout of Enterprise Frontier Safeguards later this fall. This forthcoming architecture combines zero data retention with robust guardrails, allowing eligible organizations to retain and process operational data within customer-controlled cloud environments. Currently, standard program participation mandates platform data retention for misuse monitoring, with an exception for organizations already approved for Claude Fable 5.1 or Claude Mythos 5.1 under existing zero-data-retention agreements. Organizations subject to strict data governance requirements can register interest now to prepare for deployment when Enterprise Frontier Safeguards becomes generally available.

Prospective applicants should account for differentiated review windows and ongoing classifier adjustments. Defense Access applications process within a few days, whereas Red Team Access reviews take several weeks, during which approved organizations receive interim access to Defense Access. Over coming weeks, Anthropic plans to share additional technical details on its defensive open-source partnerships and infrastructure security initiatives. Practitioners should also track ongoing refinements to tier-based classifiers across future frontier models, as well as final remediation statistics as Project Glasswing partners complete their outstanding vulnerability patch cycles.