Published: 10 October 2026
Anthropic has expanded its Cyber Verification Program, giving a larger group of qualifying cybersecurity professionals access to advanced model capabilities and reduced blocking restrictions for legitimate defensive work.
The company has folded experience from Project Glasswing into the updated programme. Glasswing brought technology companies, critical-infrastructure operators and open-source organisations together to use frontier AI for vulnerability discovery and remediation.
The problem with one set of safety controls
Cybersecurity creates a difficult access problem for AI developers. The same capabilities that help a defender find a software flaw can help an attacker discover and exploit it. Broad restrictions reduce misuse, but they can also block legitimate penetration testing, red teaming and vulnerability research.
A verified-access model tries to separate users by identity, professional purpose and accountability. Approved researchers can perform work that ordinary accounts cannot, while the provider retains monitoring and enforcement mechanisms.
Anthropic says its open-source scanning found 5,500 verified software vulnerabilities between April and October 2026. It also says partner reporting indicates broader results, although company-supplied figures should be assessed carefully because reporting is incomplete and the programme is not an independent benchmark.
Why scale changes the risk
Expanding access moves the programme from a small collaboration toward a wider operating model. That can increase defensive impact by placing capable tools with more teams. It also increases the importance of vetting, audit trails, data protection and rapid suspension when access is misused.
Security professionals often handle sensitive source code, incident evidence and details of unpatched vulnerabilities. A strong programme therefore needs clear rules governing data retention, model training, disclosure and coordination with affected software maintainers.
Reduced blocking should not mean absent control. Permissions can be limited by task, tool, target and duration, while high-risk actions may require additional human review.
A broader cyber mission
Anthropic has also announced a wider Cyber Mission that will provide engineering support, tools and funding for work protecting critical infrastructure and open-source software. This reflects a growing view that model providers must participate directly in defence, rather than only publishing general safety policies.
The initiative also raises an industry question. If access to the most capable defensive AI depends on approval by a small number of private model companies, those providers gain influence over who can conduct advanced research and under what conditions.
The experiment to watch
The success of verified cyber access should be judged through independently validated fixes, responsible disclosure outcomes, misuse rates and the time required to move from discovery to remediation.
The programme is an important test of a more mature safety model: not simply allowing or refusing a capability, but granting controlled access to known professionals for specific purposes. If it works, the approach could inform how other dual-use AI tools are governed.