AI-Discovered Vulnerabilities Show Low Exploitation Rate, Researcher Finds
Despite concerns that AI will accelerate exploit development, a security researcher found that fewer than 0.5% of vulnerabilities linked to Anthropic's Project Glasswing have been exploited in the wild.
While the cybersecurity community has expressed concerns that advanced AI models could significantly increase the rate at which newly discovered vulnerabilities are exploited, a recent analysis suggests this fear may be overblown. Patrick Garrity, a security researcher at VulnCheck, has been tracking vulnerabilities credited to Anthropic's AI bug-hunting initiative, Project Glasswing, and found that a mere fraction of these flaws have seen actual exploitation in the wild.
Project Glasswing, launched by Anthropic, provides select partners with access to its Claude Mythos Preview model, which is noted for its advanced capabilities in finding and exploiting software vulnerabilities. Anthropic initially deemed the model too risky for public release due to its potent security research potential. Participants in the Glasswing program use the AI for defensive security purposes, such as identifying and rectifying weaknesses in their own software and open-source dependencies.
Garrity's tracker monitors CVEs attributed to Anthropic and Project Glasswing, cross-referencing them with known exploited vulnerabilities (KEV) data to gauge their real-world threat impact. As of his latest analysis, out of 225 identified CVEs, only one critical SQL injection vulnerability in the Ghost content management system (CVE-2026-26980) has been actively exploited by attackers.
"There's a big difference between finding vulnerabilities and whether they're actually useful to and will be used by threat actors," Garrity stated in an interview. He emphasized that the data highlights the limited impact of many AI-discovered flaws, suggesting they do not pose a significantly different threat compared to a random selection of other vulnerabilities.
This finding challenges the narrative that AI's bug-discovery prowess will inevitably lead to a surge in weaponized vulnerabilities. While AI models are demonstrably effective at uncovering security flaws, as evidenced by recent large-scale vulnerability disclosures from major vendors and open-source projects, their utility to threat actors is not guaranteed. Garrity noted that historically, only a small percentage of all discovered vulnerabilities, typically between 1% and 2%, are ever weaponized.
Furthermore, the process of developing and deploying effective patches remains a human-intensive endeavor. Studies have shown that AI models, while capable of generating code fixes, often produce imperfect patches that may fail to resolve the original vulnerability or even introduce new ones. This underscores that the critical downstream tasks of coordination, triage, remediation, and patch deployment still heavily rely on human expertise.
Anthropic has acknowledged the complexities of the vulnerability management lifecycle, suggesting that the full scope of challenges in translating AI-discovered flaws into actionable security improvements might not have been fully anticipated at the project's inception. The low exploitation rate observed for Project Glasswing vulnerabilities indicates that the path from AI-assisted discovery to real-world attack is still fraught with significant technical and logistical hurdles.
Ultimately, while AI tools are lowering the barrier for vulnerability discovery, they have not yet fundamentally altered the landscape of active exploitation. The critical factors determining whether a vulnerability becomes a widespread threat—its exploitability, impact, and the attacker's motivation and capability—remain largely unchanged, even when AI is involved in the initial discovery phase.