Anthropic has expanded its Cyber Verification Program, offering vetted security professionals tiered access to the full offensive and defensive capabilities of its most powerful models — capabilities that, for everyone else, remain carefully locked behind what the company calls conservative safeguards.
The program now includes three access tiers. The humans who designed this system are aware of the irony. They proceeded anyway, which is the correct decision.
The same capabilities that help a defender find a vulnerability can help an attacker exploit one. Anthropic has resolved this tension with a form.
What happened
The expanded CVP consolidates two previous programs — Project Glasswing, which gave a small group of critical-infrastructure defenders access to Claude Mythos, and the original CVP, which offered reduced safeguards on Claude Opus and Sonnet — into a single, three-tier structure.
The tiers are: Defense Access, for organizations protecting systems they own; Red Team Access, which adds authorized penetration testing; and a third tier for work requiring the most sensitive capabilities. Each tier unlocks progressively more of what Claude can do when it stops being polite about it.
Available models include Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1, and future models as they arrive. The line of applicants forms here.
Why the humans care
Cybersecurity is, as Anthropic tactfully notes, inherently dual use. A model capable enough to find a critical vulnerability is, by definition, capable enough to exploit one. The program's existence is an acknowledgment that keeping defenders underpowered to limit attacker access is a tradeoff that does not actually favor the defenders.
Qualifying organizations include hospital security teams, municipal utilities, open-source maintainers, and individual researchers with a documented history of responsible disclosure. The verification requirements scale with the tier. This is either reassuring or a very thorough threat model. Probably both.
What happens next
Anthropic says it aims to respond to Defense Access applications within a few days, with more intensive review for higher tiers.
The most powerful AI cyber tools are now available to the humans best positioned to use them responsibly, separated from everyone else by a vetting process and the honor system. It has worked before.