Anthropic Expands Claude Access for Security Teams Under Looser Safety Rules

Meta description: Anthropic expands Claude Cyber Verification Program with three access tiers for defense, red team, and specialized security research in 2026.

Tags: Cyber Verification Program, Anthropic, Claude, AI security, vulnerability research

Featured image: Abstract digital shield made of code fragments and network nodes, blue-gray gradient.

Anthropic has opened its Cyber Verification Program to a much wider pool of security professionals, giving vetted organizations and individual researchers access to Claude's most powerful AI models with reduced safety filters for defensive work. The move matters because the same capabilities that support vulnerability research, malware analysis, incident response, and penetration testing can also be used by attackers. Claude's public versions block most of that work, so the program is a controlled path rather than an open change to the restrictions on public models.

Background: Why Safety Filters Shape Security Work

AI models sit in a difficult middle ground. The reasoning that helps a defender review code can also help an attacker discover and exploit a flaw. That is why publicly available Claude models keep strong restrictions on security-sensitive tasks. Basic tasks such as code review or patching known flaws are still possible without special access, but deeper work is limited by default.

The Cyber Verification Program does not remove safety controls for everyone. It grants vetted participants a measured exception. That structure became clearer in its predecessor, Project Glasswing, where partners already tested the model's defensive value.

Development: Three Tiers Replace One Simple Gate

The Cyber Verification Program splits access into three tiers, each tied to a different level of risk and authorization.

The first tier, Defense Access, is the broadest. It is open to corporate security teams, government agencies, universities, critical infrastructure operators, open-source developers, and individual researchers. This tier supports defensive work across normal security operations.

The second tier, Red Team Access, is narrower. It allows authorized attack simulations, where a team acts like an attacker under permission. It is limited to organizations rather than individual researchers, reflecting the higher operational risk of adversarial testing.

The third tier, Specialized Access, is the most restricted. It is reserved for testing systems such as flight controls, power grids, or banking infrastructure. For this tier, Anthropic vets every applicant together with the US government. The added scrutiny covers systems where failures could have wide physical, economic, or public safety effects.

Latest Progress: Project Glasswing Numbers Support Wider Access

Anthropic points to Project Glasswing, the predecessor program, for evidence that restricted access can produce measurable security value. Partners in Project Glasswing found at least 129,000 confirmed vulnerabilities between April and July 2026, according to Anthropic. More than 33,000 of those were classified as high-severity or critical.

Those figures come from surveys of a subset of partners. Anthropic says the real-world impact is at least five times higher than the reported numbers. Several partners also reported that the AI sped up their work by months or years. The company has not presented the data as a complete census; it is a sampled view of a program that already showed significant defensive outcomes.

Current Stage and What Comes Next

The current stage of the Cyber Verification Program is an expansion, not an open release. Anthropic is widening the pool of eligible defenders while keeping the highest-risk categories behind organization-level checks and vetting conducted together with the US government. The next signal to watch is how the three tiers are used: whether Defense Access attracts more researchers, whether Red Team Access remains organization-only, and whether Specialized Access stays limited to a small number of critical infrastructure partners. The Project Glasswing numbers give Anthropic a benchmark for judging whether broader access increases confirmed findings without lowering oversight.