Seven months into Project Glasswing, Anthropic’s flagship AI vulnerability research initiative, the numbers tell a sobering story. VulnCheck has tracked 442 CVEs attributable to the project — roughly 0.5% of all CVEs published this year — with only two identified as exploited in the wild.
The real problem is not the low hit rate. It is that duplicates are overtaking valid fixes, placing an unexpected burden on the open-source maintainers the project was supposed to help. When an AI floods maintainers with marginal findings that another tool already flagged, the net effect is noise, not security.
The gap between promise and practice
Glasswing was announced with the kind of ambition that AI security research often attracts: automated vulnerability discovery at scale, reducing the attack surface of critical infrastructure. What has emerged instead is a validation workflow that produces more work for humans than it saves. The ratio of duplicates to genuine finds has grown steadily, and most reported CVEs remain unexploited — meaning the tool is efficient at generating paper, not at catching real threats.
The Qur’anic reminder that it is despicable to say what one does not do (61:3) applies here in a specific, non-moralising sense. The gap between what the project promised and what it delivers isn’t a failure of intention; it is a failure of measurement. When an AI security tool’s primary output is noise that distracts overworked maintainers, the real vulnerability is not in the code — it is in the incentive structure that rewards announcements over outcomes.