Anthropic just consolidated Project Glasswing into one unified Cyber Verification Program with tiered access to Opus 5.5, Sonnet 5.5, and Mythos 5.1.
Here's what matters:
→ Three tiers: Defense, Red Team, Specialized—each with custom safeguards
→ Open-source maintainers and vetted researchers can now apply for access
→ Glasswing partners crushed it: 129,000+ verified vulnerabilities found in just 4 months
This is a massive play for AI security infrastructure. If you're building in the security or AI tooling space, this framework could set the standard for how models get stress-tested at scale.
Also worth noting: giving researchers structured access to frontier models accelerates both offense and defense. The vulnerability discovery rate is insane—expect more disclosures, more patches, and potentially more exploits in the wild as this scales.
Here's what matters:
→ Three tiers: Defense, Red Team, Specialized—each with custom safeguards
→ Open-source maintainers and vetted researchers can now apply for access
→ Glasswing partners crushed it: 129,000+ verified vulnerabilities found in just 4 months
This is a massive play for AI security infrastructure. If you're building in the security or AI tooling space, this framework could set the standard for how models get stress-tested at scale.
Also worth noting: giving researchers structured access to frontier models accelerates both offense and defense. The vulnerability discovery rate is insane—expect more disclosures, more patches, and potentially more exploits in the wild as this scales.