Prompt injection ranks No.1 on OWASP, No.12 in real incidents

The OWASP Top 10 for LLM Applications has listed prompt injection as the top risk for three years, reflecting its theoretical severity. But when two project leaders checked the list against real-world incidents, they found it appears less frequently in labeled events—at No. 12. This does not mean prompt injection is less dangerous; rather, it suggests that current vulnerability scanning methods often miss it, so it goes underreported in incident databases.

The study's data comes from multiple sources, including CVE and the AIAAIC AI-harm database, covering 7,714 incidents, of which 6,639 were labeled. Because the analysis is exploratory and not official, its findings should be seen as a signal for deeper investigation rather than a revision of the OWASP list. For security leaders, the key takeaway is that low CVE counts are not a reliable measure of risk for prompt injection.