OWASP LLM Top 10

What it means

The OWASP Top 10 for Large Language Model Applications is a ranked list of the most important security risks in software built on language models. It comes from the OWASP GenAI Security Project, a community effort, and has been revised since the first edition as the field has changed.

The current list includes prompt injection, sensitive information disclosure, supply chain risk, data and model poisoning, improper output handling, excessive agency, system prompt leakage, weaknesses in vector and embedding stores, misinformation and unbounded consumption. Each entry describes the risk, examples and ways to reduce it.

Why it matters

It gives a team a shared vocabulary. A product manager asking whether the assistant can leak another customer’s data, or an engineer deciding how much the agent may do, can point to a named risk rather than a vague worry. Buyers also use it as a reference when they review AI features.

It is useful because the risks are mostly familiar problems in a new setting. Weak access control still lets one account see another’s data, whether the route goes through a form or a chat prompt. Unsafe output still becomes a browser script. The list helps spot where the old problems have moved.

How it shows up

Use it as a review checklist for each AI feature. Start with prompt injection and excessive agency, because they decide what an attacker can do. Then check where the model reads data from, where its output goes next, and what it costs when someone sends it huge inputs repeatedly.

Mistakes repeat. Teams test the prompt and forget the API behind it. Retrieval pulls documents without checking who may see them, which is a RAG security gap. Output goes straight into a page or a query. Treat the list as a starting point and test the whole application, not the model alone.