AI Finds Thousands of Vulnerabilities — Only a Handful Get Exploited

VulnCheck analyzed 1,061 vulnerabilities discovered by AI in the first half of 2026: only 14 of them (1.3%) were actually exploited in attacks. Anthropic's Glasswing project found 23,000 issues — just one was targeted. The same exploitation rate as without AI at all.

The paradox is that AI is great at finding holes but drowns defenders in thousands of findings — separating the dangerous from the background noise keeps getting harder. Meanwhile, AI systems themselves, agents and model interfaces, are becoming a new attack surface that the industry has yet to monitor in any systematic way.