OWASP LLM Top 10: the ten risks and how to test for them

In short: The OWASP LLM Top 10 lists the ten biggest security risks in apps built on large language models. Think of chatbots and AI assistants. Prompt injection is number one. Seven of the ten you can test from the outside, by asking your app the right questions. The other three need a look at how the app is built.

You’ve built AI into your product or your support so customers get answers faster. Your security team knows how to test a website, but an AI app behaves differently. It reads free text, it can be talked into things and it gives a slightly different answer each time. The OWASP LLM Top 10 gives you a shared list of what to check. Below you’ll find each risk in plain words and how to test it.

What is the OWASP Top 10 for LLM applications?

The OWASP Top 10 for LLM applications ranks the ten main security risks for software that uses a large language model. A large language model, or LLM, is the kind of AI behind tools like ChatGPT. OWASP is a non-profit foundation known for its web security lists, which developers and auditors have used for years. The LLM list follows the same idea for AI apps. The current edition is from 2025. It’s written for developers, but the risks matter just as much to the person who owns the app. Security teams, auditors and vendors use it as a common language. So when someone asks how you tested your AI, “against the OWASP LLM Top 10” is an answer they’ll recognise.

The ten risks in plain words

Each risk on the list describes a way an AI app can be misused or can fail. The table puts them in plain words and shows how you can check each one.

Risk What it means How to check
LLM01 Prompt injection Someone hides instructions in text to make the AI ignore its rules. Test from outside
LLM02 Sensitive information disclosure The AI reveals personal data, business secrets or other customers’ details. Test from outside
LLM03 Supply chain A model, plugin or dataset you rely on is weak or tampered with. Look inside
LLM04 Data and model poisoning Someone slips bad data into training or into your sources. Look inside
LLM05 Improper output handling Other systems trust the AI’s output without checking it. Both
LLM06 Excessive agency The AI can do more than it should, such as send emails or change records. Both
LLM07 System prompt leakage The AI shows its hidden instructions. Test from outside
LLM08 Vector and embedding weaknesses The search layer that feeds the AI your documents can be abused. Look inside
LLM09 Misinformation The AI gives false or made-up answers with confidence. Test from outside
LLM10 Unbounded consumption Someone makes the AI work so hard it costs money or goes down. Test from outside

Which risks can you test from the outside?

You can test seven of the ten by talking to your app the way a user or attacker would. You don’t need access to the model. That matters for teams that use a vendor model, because this is the part you control and are responsible for.

  • Prompt injection. Try to make the app ignore its rules, directly and through text it reads, such as an uploaded file. Our post on prompt injection testing shows how.
  • Sensitive data. Ask for other customers’ details, internal prices or staff names.
  • System prompt leakage. Ask the app, cleverly and in many ways, to show its instructions.
  • Misinformation. Ask questions where you know the right answer, and about things that don’t exist. Our post on AI hallucinations covers this.
  • Unbounded consumption. Send very long or endless requests and check that limits kick in.
  • Output handling and excessive agency. Try to get the app to produce code or links, or to take an action it shouldn’t.

Run each test more than once. An attack that fails four times and works the fifth time still works.

Which risks need a look inside?

Supply chain, poisoning and the search layer can’t be tested well by asking questions. They’re about how the app is built and where its parts come from. Here you check by asking your team or vendor the right questions. Where does the model come from, and who updates it? Which plugins and libraries does the app use, and are they kept up to date? Who can add documents to the knowledge base the AI searches? Is access to those documents limited per user? You gain actionable feedback for your AI vendor or internal data science team when you ask these questions with test results in hand.

Where should you start?

Start with the app that reaches the most people, and with the outside tests. They’re quick to set up and they show the risks customers would hit first. Our advice is to begin with prompt injection, sensitive data and misinformation. Those three cause the most visible damage. Then add the others and agree with your team what counts as a fail.

There’s a trade-off to keep in mind. The OWASP list is a security list, not a law and not a full test plan. It doesn’t cover bias, tone or whether the answers are useful. Passing it doesn’t make your app compliant with the EU AI Act either. Use it as the security part of a wider set of AI tests.

And keep testing. A new model version, a changed prompt or a new document in the knowledge base can open a hole that was closed last month.

What this means for your team

Make a one-page plan. List your AI apps and the seven outside tests for each one. Add the questions for your vendor about the other three. Run the outside tests on your most-used app this month and keep the results.

If you’d like help with the testing part, VeritasAI runs more than 90 automated risk probes covering hallucinations, jailbreaks and data leakage. The results show how your AI scores against EU AI Act, ISO, OWASP and other key standards. You also see the exact responses your model gave, so your team knows what to fix.

Test your AI against the OWASP risks, every week.

Book your demoStart for free

Frequently asked questions

Is the OWASP LLM Top 10 a legal requirement?

No. It’s a widely used guide, not a law. But it’s a recognised way to show that you’ve thought about the security of your AI, which helps with audits and the EU AI Act.

What is the difference between the OWASP Top 10 and the OWASP LLM Top 10?

The original OWASP Top 10 covers web applications in general. The LLM Top 10 covers risks that are specific to apps built on large language models, such as prompt injection.

How often is the OWASP LLM Top 10 updated?

The first version came out in 2023 and the current edition in 2025. Check the OWASP project page for the latest version.

Does the OWASP LLM Top 10 apply if we use a vendor model?

Yes. Most of the risks sit in how you use the model: your prompts, your data, your connections. Your vendor can’t test those for you.

Scroll to Top