What is the OWASP Top 10 for LLMs?
The OWASP Top 10 for LLMs, told as a story about an industry-wide board that gave shared numbers to incidents every station used to call by a different name.
An unfamiliar board appeared at the entrance.This board isn't just our own records — it's a list the whole industry put together.
Every station calls the same incident by a different name.When the same incident has different names, stations can't compare notes or learn together.
The OWASP LLM Top 10 is a shared, industry-agreed risk list.Each number is just a label — the detail lives elsewhere.
Each Top 10 item links out to its own fuller story.This page explains the list itself — the details of each item live on the linked page.
The board itself doesn't block any incident.The list is only a map — the actual defenses are on each item's page.
The OWASP Top 10 for LLMs = a shared risk board that pools incidents from many sites into industry-agreed numbers. The list itself is only a map — the real defenses live on the page each number points to.
The OWASP Top 10 for LLM Applications is an LLM01-LLM10 threat classification that the OWASP GenAI Security Project updates periodically from community judgment and analysis of real incidents. In the 2026 edition, published in August 2026, its items in order are prompt injection, sensitive information disclosure, excessive agency, supply chain, data and model poisoning, unbounded consumption, misinformation, hidden context exposure (a broader version of the 2025 edition's system prompt leakage), vector and embedding weaknesses, and improper output handling. Numbers shift between editions, so cite the edition with the number. The list's value is in giving everyone a shared vocabulary — the actual defenses live in the detailed material (guidelines, or this site's own per-term pages) that each item points to.
When grown-ups say it
- LLM01 Prompt Injection
- The note slipped behind the order. → see the note incident in full
- LLM05 Data and Model Poisoning
- Tainted crops mixed into the pantry. → see the tainted-ingredient story
- LLM08 Hidden Context Exposure
- Quietly copying hidden instructions, like the recipe book. → see the leak in full
- LLM03 Excessive Agency
- An errand given too much freedom. → see it in full
- Model Theft
- Copying the whole sorting rulebook. LLM10 in the 2023 edition; the 2026 edition folds it into LLM06, unbounded consumption. → see it in full
- AI Governance
- The whole-farm operating rules board. The Top 10 is one list governance leans on. → the rules board on the wall
- Red Team
- Someone who actually pokes at each board item, unannounced. → the AI-specific inspector
- Model Card
- The spec sheet for one sorter. Safety check results are sometimes noted there too. → see it in full