OWASP Top 10 for LLM Applications 2026
OWASP's ten top security risks for applications built on large language models. The risks, LLM01 to LLM10, run from prompt injection and sensitive information disclosure to vector and embedding weaknesses and improper output handling. Each comes with attack scenarios and guidance on preventing it. This edition renumbers eight of the ten risks of 2025 and widens System Prompt Leakage into Hidden Context Exposure. OSA's pattern SP-047, Secure Agentic AI Frameworks, cites Excessive Agency, which is LLM03 here and was LLM06 in 2025.
How this mapping was made
OSA's own analysis, from OWASP's text by three independent ratings. NIST's catalogue of registered crosswalks has none from this list to SP 800-53 (checked 1 November 2026). The OWASP GenAI Security Project also keeps a crosswalk on GitHub that names SP 800-53 controls for each risk. Each of the ten risks was rated three times, independently, by AI models working from OWASP's text and the control statements. No rater saw another's rating. A control is listed for a risk where at least two of the three ratings named it. The coverage figure is the median of the three estimates of how much of the risk, and of OWASP's guidance for preventing it, the listed controls express. The rationale and gaps were written after the controls were settled, from the raters' reasons and from the gaps at least two of them named.
Texts
- OWASP Top 10 for LLM Applications 2026 (OWASP GenAI Security Project, Version 2026, August 2026)
| Clause | Title | SP 800-53 Controls |
|---|---|---|
| LLM01 | Prompt Injection | |
| LLM02 | Sensitive Information Disclosure | |
| LLM03 | Excessive Agency | |
| LLM04 | Supply Chain | |
| LLM05 | Data and Model Poisoning | |
| LLM06 | Unbounded Consumption | |
| LLM07 | Misinformation | |
| LLM08 | Hidden Context Exposure | |
| LLM09 | Vector and Embedding Weaknesses | |
| LLM10 | Improper Output Handling |
Clause titles are the source's own wording, from OWASP Top 10 for LLM Applications 2026, by the OWASP GenAI Security Project (August 2026), used unchanged under CC BY-SA 4.0.