A researcher tracking vulnerabilities tied to Anthropic's Project Glasswing initiative found that only one of 225 flaws has been confirmed exploited in active attacks, according to data maintained by VulnCheck security analyst Patrick Garrity. The findings emerge despite Anthropic's emphasis on the advanced bug-finding capabilities of its AI models, which were made available exclusively to vetted partners for defensive security research following the program's April launch. Garrity's analysis cross-references these CVEs with known exploited vulnerabilities to assess real-world threat relevance. He noted that while AI systems can identify numerous security gaps, their utility to attackers remains limited. 'There's a big difference between finding vulnerabilities and whether they're actually useful to and will be used by threat actors,' Garrity told The Register. Historical data shows only about one to two percent of discovered vulnerabilities typically enter exploitation campaigns. Recent studies reinforce this pattern: AI-generated code fixes successfully resolved vulnerabilities in just 26 percent of cases across 6,080 tested patches, with over half introducing new issues or failing entirely. Similarly, Veracode reported an average 56 percent security pass rate for AI-generated code across multiple models and tasks. Garrity emphasized that while AI lowers the barrier to vulnerability discovery, downstream processes like patch deployment and triage remain predominantly human-driven. Anthropic has not yet responded to inquiries about these findings.