Enterprise AI in Numbers · Risk and regulation
New prompt injection vulnerabilities published each quarter
49
49 as of 2026 Q3, according to NIST National Vulnerability Database, up from 31 in 2026 Q2. CVE records published in the quarter whose description mentions prompt injection.
What it means for leaders
Prompt injection has moved from research papers to catalogued vulnerabilities in shipping products. Treat any AI feature that reads untrusted content (email, web pages, documents) as an attack surface, with least-privilege tools and output checks.
What would change the reading: Quarter-over-quarter growth, and vulnerabilities in agent frameworks and connectors.
The trend
Every reading
Definition and method
- What it measures
- CVE records published in the quarter whose description mentions prompt injection.
- Population
- CVE records in the NIST National Vulnerability Database
- Updated
- Quarterly
- Source
- NIST National Vulnerability Database · Public domain (US government work)
Caveats
- Keyword matching misses records that describe the flaw in other words.
- Disclosure practices for AI flaws are still forming, so counts understate the problem.
Cite this
Enterprise AI in Numbers, "New prompt injection vulnerabilities published each quarter", https://lucianoayres.com/ai-brief/data/llm-vulnerabilities/, retrieved October 25, 2026. Data: NIST National Vulnerability Database.