Anthropic Opens Its Strongest Claude Models: Cybersecurity Program Gets Three-Tier White-Hat Access

On October 6, Anthropic announced a revamped Cyber Verification Program (CVP). Vetted cybersecurity professionals can now use its strongest models — Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 — with fewer safety restrictions to find vulnerabilities.

The decision comes with a track record. Glasswing partners found at least 129,000 verified vulnerabilities between April and July. Anthropic’s own open-source scanning turned up another 5,500 between April and October. More than 33,000 were rated critical or severe.

Glasswing targets the world’s most critical software, working with teams that secure critical infrastructure and major open-source projects. In other words, those 129,000 vulnerabilities weren’t generic scan results — the search started with the software that matters most. Anthropic says the number may be an underestimate, since only some partners were surveyed and the real impact could be 5x.

Two Old Programs, Merged Into One

CVP isn’t starting from zero. It merges two programs Anthropic has run over the past six months.

One is Glasswing: giving organizations that secure critical software access to Claude Mythos, Anthropic’s most cyber-capable model family. The other is the original CVP: relaxing Opus and Sonnet safety restrictions for vetted security teams.

After the merge, there are three access tiers: Defense, Red Team, and Specialized.

What does “fewer safety restrictions” actually mean? Standard Claude typically refuses or truncates requests about exploit details and attack techniques. The CVP version relaxes those limits — on the condition that it’s used only for legitimate defensive work. That’s the gate the whole program turns on: the tool’s permissions are tied to its purpose.

How the Three Tiers Work

Defense covers defensive work: incident response and malware analysis. Security teams, critical infrastructure operators, open-source project maintainers, and researchers with a vulnerability-reporting record can all apply.

Red Team goes one step further: authorized penetration testing and red-team exercises. This tier only accepts organizations — individuals need not apply.

Specialized is the least restricted tier, open to a handful of organizations. It’s for testing safety-critical systems like power grids, flight systems, and interbank transfer systems.

All three tiers get Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1, plus future models. Members are vetted jointly by Anthropic and the U.S. government; existing Glasswing members transfer over directly.

Why Now

When Claude Mythos Preview first appeared in April, the market’s worry was that AI would be used to hack software first: finding vulnerabilities faster than they’re patched. Anthropic’s approach is the reverse: get the tools into white hats’ hands first, so the defense starts running ahead.

The numbers say the approach has worked: four months, 129,000 vulnerabilities, more than 33,000 critical or severe. This revamp turns the pilot into a formal three-tier program with written admission standards.

One detail worth noting: member vetting is done jointly by Anthropic and the U.S. government. In other words, this isn’t just a business partnership — it’s inside the government’s security cooperation framework. The Specialized tier can touch power grids, flight systems, and bank transfer systems; who holds the vetting power decides who gets that key.

Two things to watch next: which organizations end up approved for the Specialized tier, and how the real offensive-defensive gap plays out for these models. At least this year, in vulnerability hunting, AI became the defender’s tool first.