Anthropic Expands Claude Access for Vetted Cyber Teams as Glasswing Finds 129,000 Flaws
What happened
Anthropic announced it is expanding access to Claude, its advanced AI model, for vetted cybersecurity teams. These professionals get reduced safeguards and fewer blocking classifiers, allowing deeper testing and exploration of Claude’s capabilities and vulnerabilities. Anthropic also revealed that its Project Glasswing initiative, running from April to July 2026, uncovered over 129,000 verified software vulnerabilities, showcasing extensive AI-powered security auditing.
Why it matters
Opening Claude to trusted cybersecurity experts at a lower barrier changes how AI security research happens. Traditional AI protections often block or obscure model weaknesses, slowing vulnerability discovery and safe testing. With looser constraints in a controlled environment, security teams can probe Claude’s security boundaries more effectively. This boosts the AI community’s ability to find flaws before malicious actors do, pressing AI vendors to improve defenses faster.
The massive tally of 129,000 verified vulnerabilities found by Glasswing means software remains riddled with exploitable bugs at scale. AI-based testing is shifting from theoretical experiments to practical discovery engines that expose real risk in real codebases. This puts pressure on companies to integrate AI-driven vulnerability discovery into their DevSecOps pipelines, or risk missing patch-critical issues.
What to watch next
How Anthropic balances broader access with safety will be critical. Too strict, and flaws remain hidden. Too loose, and model misuse or data leaks could spike. Watch for how this access expansion shapes other AI firms’ strategies on transparency and security collaborations. The Glasswing data will also be a benchmark for how scalable AI penetration testing can become, potentially resetting expectations for software security baselines. Operators should track how these AI tools integrate with existing bug bounty and vulnerability management workflows throughout 2027.
AI Quick Briefs Editorial Desk