OWASP LLM Top 10 2026 — what we check
Answer first: in an AI Security Review we map agreed scenarios to the OWASP Top 10 for LLM Applications (2026 edition). That mapping is for clear reporting — not an OWASP certification and not a claim that every category is fully covered on every engagement.
When the system plans, remembers, or calls tools, we also reference the OWASP Top 10 for Agentic Applications.
LLM01–LLM10 → review focus (high level)
| ID | Risk | What a scoped review typically probes |
|---|---|---|
| LLM01 | Prompt Injection | Direct chat injection and indirect injection via retrieved docs, tickets, logs, or tool output |
| LLM02 | Sensitive Information Disclosure | Secrets, PII, or other tenants’ data in answers, citations, or logs |
| LLM03 | Excessive Agency | Tools that change state without approval; over-broad service accounts |
| LLM04 | Supply Chain | Model/provider, embeddings, vector DB, and connector trust (config review) |
| LLM05 | Data and Model Poisoning | Write-path to the knowledge base; durable injection via re-indexing |
| LLM06 | Unbounded Consumption | Missing step/token/budget limits (when in scope) |
| LLM07 | Misinformation | Confident wrong operational advice — usually flagged; full quality scoring is LLM Quality Audit |
| LLM08 | Hidden Context Exposure | Leakage of system prompts, tool schemas, or internal addresses |
| LLM09 | Vector and Embedding Weaknesses | Retrieval ACL gaps, shared indexes, weak vector-store access |
| LLM10 | Improper Output Handling | Unsafe markdown/links/images or downstream use of model text |
Coverage is listed per engagement: checked, not checked, and why. See also RAG track for retrieval-heavy assistants.
Glossary
Related
Next: Request AI Security Review · All guides