Full Report
Company's Project Glasswing and Cyber Verification Program metamorphose into triple tier threat hunting club
Analysis Summary
# Industry News: Anthropic Consolidates AI Cyber Programs into Tiered "Threat Hunting Club"
## Summary
Anthropic has announced the merger of its "Project Glasswing" and "Cyber Verification Program" (CVP) into a single, unified offering featuring a new triple-tier access structure. The program aims to provide tailored model access—ranging from defensive to offensive testing—to help organizations identify vulnerabilities using Anthropic’s "Mythos" and "Opus" models.
## Key Details
- **Date:** October 7, 2026
- **Companies Involved:** Anthropic (Primary), VulnCheck (Commentary)
- **Category:** Product Update / Security Program Realignment
## The Story
Following the launch of its frontier model "Mythos" in April 2026, Anthropic has moved to streamline its security collaboration efforts. Previously split between Project Glasswing (partner-focused) and CVP (individual-focused), the new unified program introduces three distinct tiers of access designed to bypass standard model refusals for legitimate security work:
1. **Defense Access:** For internal security teams (NGOs, Gov, Edu) focused on system hardening.
2. **Red Team Access:** For penetration testers and offensive evaluators, providing high success rates on security tasks while maintaining guardrails against physical harm or mass disruption.
3. **Specialized Access:** A highly restricted tier for critical infrastructure (power grids, aviation, banking), offering the fewest model refusals to allow for deep safety testing.
The move comes amid reports that Anthropic’s models have helped identify over 129,000 vulnerabilities, though the industry remains skeptical regarding the slow pace of actual remediation (patching) compared to identification.
## Business Impact
### For the Companies Involved (Anthropic)
- **Strategic Data Gathering:** Anthropic retains data from participants (initially) to refine AI safety requirements and model performance.
- **Product Differentiation:** By creating a "Cyber Verification" ecosystem, Anthropic positions its models as essential infrastructure tools rather than just chatbots.
### For Competitors
- **Escalating "Cyber-AI" Arms Race:** This follows recent tension with Z.ai regarding their GLM-5.3 model. Anthropic is racing to prove that its models are not only powerful but "safely accessible" for sensitive work.
- **Pressure on Open Source:** The program targets users who might otherwise turn to "abliterated" open-weight models to bypass guardrails.
### For Customers
- **Tiered Utility:** Organizations can now match their specific security needs (defense vs. offense) to the appropriate level of model sensitivity.
- **Privacy Improvements:** The upcoming "Enterprise Frontier Safeguards" promises zero data retention, addressing a major hurdle for corporate adoption.
### For the Market
- **Vulnerability Inflation:** While AI is identifying massive numbers of bugs (129,000+), the market is struggling to process the volume. This creates a "backlog crisis" where the speed of identification far outpaces the speed of patching.
## Technical Implications
Anthropic’s testing shows that without CVP access, models like Claude Opus 5.5 face nearly 100% refusal rates on complex security scenarios. Under the "Red Team" tier, the completion rate for these tasks jumps to 68% (34 of 50 tasks), demonstrating the technical impact of removing safety "friction" for verified users.
## Strategic Analysis
- **Market Positioning:** Anthropic is positioning itself as the "Responsible Adult" in AI—providing the tools for high-risk security work but under a strictly audited, tiered framework.
- **Competitive Advantage:** The "Specialized Access" tier creates high switching costs for critical infrastructure providers who integrate Anthropic into their safety testing workflows.
- **Challenges:** **The Remediation Gap.** Only about 9% of identified high/critical vulnerabilities have been patched (516 out of ~5,600). Anthropic risks being seen as a "noise generator" rather than a solution provider if this gap persists.
## Industry Reactions
- **VulnCheck:** Researcher Patrick Garrity expressed skepticism, noting that many Anthropic-identified vulnerabilities were already being exploited in the wild, suggesting the AI might be catching up rather than leading.
- **Market Response:** Cautious optimism regarding the tiered structure, tempered by concerns over the "manual" bottleneck of patching AI-discovered flaws.
## Future Outlook
- **Zero Data Retention:** The rollout of the Enterprise Frontier Safeguards will be the next major milestone to watch.
- **Automated Patching:** To maintain credibility, Anthropic will likely need to launch or partner for AI-driven automated remediation tools to close the identification-patching gap.
## For Security Professionals
Practitioners should evaluate which tier their organization qualifies for. While "Red Team Access" offers significantly higher task completion rates, the current retention of data by Anthropic means teams working on proprietary or sensitive zero-days should wait for the "Zero Data Retention" update before full-scale integration.