Anthropic Opens Stronger Cyber AI to More Defenders
Matthew Leo · Published October 9, 2026 · News
Anthropic is widening access to its most capable cybersecurity models, but it is not simply switching off safeguards for anyone who asks. Its expanded Cyber Verification Program uses three tiers, identity checks and different controls for defensive work, authorized red teaming and testing of high-risk systems.
The October 6 announcement matters to Canadian security teams because the entry tier is broad enough to include companies, universities, governments, smaller security firms, open-source maintainers and individual researchers with a record of responsible disclosure. Access still depends on Anthropic approving the applicant.
What the three access tiers allow
Defense Access covers work such as incident response, malware analysis and validating vulnerabilities in systems the applicant owns or maintains. Anthropic says many legitimate defenders should qualify and that it aims to respond within days.
Red Team Access adds authorized penetration testing. It is limited to organizations, and applicants must show that the systems they test are within their authority. Anthropic says review can take several weeks. The company also says actions that could cause physical harm or mass disruption, including ransomware deployment, remain blocked.
Specialized Access has the fewest cyber restrictions. It is reserved for a small group authorized to test systems such as power grids, telecom networks, flight operations and interbank infrastructure. Anthropic says it reviews those organizations with the U.S. government. That creates an obvious diligence question for Canadian critical-infrastructure operators: what information is shared during approval, monitoring and incident handling?
The results are promising, but they are company-reported
Anthropic says partners using its earlier Project Glasswing program found at least 129,000 verified software vulnerabilities between April and July 2026, including more than 33,000 rated critical or high severity. Its own open-source scans found another 5,500 verified vulnerabilities between April and October.
Those figures are not an independent audit. Anthropic says they come from partial partner reports using different triage methods, and fewer than half of partners disclosed patch counts. The numbers show scale, but they do not tell readers how many findings were duplicates, how severity was validated across partners or how many issues were fixed.
The company also tested Claude Opus 5.5 on ten multi-stage cyber challenges, with five attempts per challenge. Its Defense setting blocked 46 of 50 trials at some point. In Red Team mode, no blocks occurred and the model completed 34 of 50 trials. That result illustrates how much capability the verification gate controls; it is not proof that a team will achieve the same performance on its own environment.
What a Canadian applicant should settle first
- Define authorization. List the systems, subsidiaries, clients and cloud accounts the team may test. A customer contract should state the testing scope rather than relying on a general services clause.
- Separate access by workspace. Anthropic requires administrators to assign program access to specific workspaces. Keep ordinary development, defensive operations and red-team work apart, with named owners and activity logs.
- Review data handling. Anthropic says data retention is required for most program participants while it monitors misuse. Teams handling Canadian personal information, health data or regulated records should confirm retention, storage location, subcontractors and deletion terms before uploading evidence.
- Plan human validation. A model finding is a lead. Reproduce it safely, remove duplicates, assign severity consistently and coordinate disclosure before treating it as verified.
- Protect credentials. Give the testing environment only the secrets and network reach it needs. A powerful defensive model does not reduce the impact of an exposed production key.
Teams that only need code review, patching known issues or vulnerability finding in source they own may not need the program at all. Anthropic says those tasks remain available through its generally offered models.
The Canadian opportunity and the limit
Hospitals, municipal utilities, universities and open-source maintainers often have more software to review than people available to review it. Faster triage could help, especially when a newly disclosed flaw affects many systems. The value comes from shortening the path from inventory to validated fix, not from producing the largest possible pile of findings.
Mapletechie recently covered both a critical self-hosted GitLab AI Gateway flaw and attack-style probes sent to Library and Archives Canada. Together, they show why capability and authorization must be addressed at the same time. Better tools can help defenders, but they also make boundaries, logging and review more important.
Sources
- Anthropic: expanded Cyber Verification Program, tiers and reported results
- Reuters: independent reporting on the program expansion
Tags: Anthropic, Claude, Cyber Verification Program, cybersecurity, Canada