AI Finds Thousands of Vulnerabilities — Only a Handful Get Exploited

VulnCheck examined 1,061 vulnerabilities identified through AI during H1 2026. Of those, 14 (1.3%) were used in attacks. The Glasswing project from Anthropic identified 23,000 vulnerabilities. One was exploited. That’s the same rate of exploitation as if there were no AI involved.

It turns out that while AI is fantastic at spotting flaws, it generates thousands of results that make it difficult to distinguish between what’s important and irrelevant. At the same time, AI-powered tools, agents, and models present a new attack surface which we aren’t currently keeping track of.