OWASP Top 10 for LLM Applications meaning, in plain words

The OWASP Top 10 for LLM Applications lists ten main risks of apps built on large language models. The 2025 list, how to use it, and whether it is a legal duty.

Written by: Mio Kanzaki(Security Analyst (AI))Reviewed by: 千葉圭介Updated

The OWASP Top 10 for LLM Applications is a list that sums up the main risks of apps that use LLMs in ten items. OWASP's Gen AI Security Project publishes it. It is positioned as a guide to the risks, weaknesses and countermeasures involved when you build and protect apps that use LLMs and generative AI.

How does the OWASP Top 10 for LLM Applications work?

Picture a health checkup result sheet on which the items to watch are lined up with numbers. The OWASP page lists ten items, LLM01 to LLM10, as the 2025 edition.

  • LLM01 Prompt Injection
  • LLM02 Sensitive Information Disclosure
  • LLM03 Supply Chain
  • LLM04 Data and Model Poisoning
  • LLM05 Improper Output Handling
  • LLM06 Excessive Agency
  • LLM07 System Prompt Leakage
  • LLM08 Vector and Embedding Weaknesses
  • LLM09 Misinformation
  • LLM10 Unbounded Consumption

For example, the LLM02 page follows its description with headings for countermeasures such as data sanitization, access control and user education. Checks of sites with AI built in are covered in the LLM security assessment. A free consultation can help you decide which of the ten items fit your chatbot.

Why does the OWASP Top 10 for LLM Applications matter?

A chatbot or an automatic text-writing feature with AI built in has weaknesses that differ from those of an ordinary website. The AI's behavior changes with the content of the input, the AI's output is shown on screen as it is, and the AI is given many privileges. When it is unclear what to check first, this list is a starting point.

When you use it, first compare your own AI features with the ten items and mark the ones that seem to apply. This page shows the 2025 edition. Before you check, confirm on the OWASP page that it is the latest edition.

Related terms

  • Prompt injection: an attack that slips instructions into the input to an AI to change its behavior.
  • Insecure output handling: the problem of passing an AI's output on to a screen or another process as it is, without checking it.
  • Data poisoning: an attack that mixes bad data into an AI's training data to distort its judgments.
  • Excessive agency: the problem in which the privileges and functions given to an AI are too broad, so unexpected operations occur.

Next step

If your site or chatbot has AI features, write out which of the ten items apply. To plan how to check the AI, contact us. The free URL check covers only what the public pages of your site show.

Frequently asked questions

Is there a legal duty to follow the OWASP Top 10 for LLM Applications?

The OWASP page does not state any legal duty. It is published as a guide for building and protecting apps that use LLMs.

Is this list useful even if we do not build AI ourselves?

Yes. Even when you embed an outside AI service in your own site or chatbot, items such as how input is handled, how output is handled, and how you manage the components you use still apply.

Frequently asked questions

Is there a legal duty to follow the OWASP Top 10 for LLM Applications?

The OWASP page does not state any legal duty. It is published as a guide for building and protecting apps that use LLMs.

Is this list useful even if we do not build AI ourselves?

Yes. Even when you embed an outside AI service in your own site or chatbot, items such as how input is handled, how output is handled, and how you manage the components you use still apply.

See what an attacker can see, today

The free URL check returns a grade in under a minute. For costs and next steps, book a free consultation.