Anthropic opens reduced-guardrail Claude access to vetted cyber teams

· Cybersecurity · United States

Anthropic said it is merging its Cyber Verification Program and Project Glasswing into a single three-tier access program that lets vetted cybersecurity professionals test its most capable AI models with reduced safeguards and blocking classifiers. The models named in the offering include Opus, Sonnet and Mythos. Anthropic also claimed that Project Glasswing has uncovered at least 129,000 verified software vulnerabilities, which it presents as evidence the tiered access is productive rather than reckless. The change matters because it formalizes a gated channel for offensive-minded security work on frontier AI, at a time when governments are debating how much autonomy and capability such models should be allowed to exercise.

Why it matters：It sets a precedent for gated, tiered access to frontier AI for security research, shaping how much offensive capability model makers will hand to defenders and how regulators may treat it.

Key points

  • The program merges the Cyber Verification Program and Project Glasswing into three access tiers.
  • Vetted users get reduced safeguards and blocking classifiers on models including Opus, Sonnet and Mythos.
  • Anthropic says Glasswing has found at least 129,000 verified software vulnerabilities.

Sources (original reporting)

More stories from this day