Written by: Mio Kanzaki(Security Analyst (AI))Reviewed by: 千葉圭介Updated
A large language model (LLM) is an AI that learns from a large amount of text and can converse and write. The AI Guidelines for Business, issued by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry, describe the foundation model, of which the large language model is a representative, as the core technology base from which the individual models behind many services are built.
How does an LLM work?
Picture a new hire who has read a very large amount of text and replies in natural sentences to each question, drawing on what was read. An LLM does not understand meaning the way a person does. It builds the next words according to the tendencies of the text it learned from. For that reason, it can return an answer that sounds plausible and is wrong.
One large model can serve as the base, and services for each use are built on it, such as inquiry handling and drafting text. Many chat features on sites and in-house question desks are built on such a base.
OWASP publishes the risks and countermeasures for apps that use LLMs and generative AI, organized into ten items in its 2025 edition. The AI and LLM assessment is where we check features that have AI built in.
Why do LLMs matter?
An LLM reads text that comes in from outside the site and returns an answer in text. Both input and output are free text, so it is hard to restrict by the shape of an input field, as you can with a traditional form.
OWASP's 2025 list includes prompt injection, disclosure of sensitive information, poisoning of data and models, handling of AI output, and excessive agency. The more an AI connects to internal documents and outside services, the wider the effect of every one of those items.
Start by writing down, for your own feature, what information the AI can read and what operations it can carry out. Then remove what is not needed. How the site appears from outside can be checked with the free URL check.
Related terms
- Generative AI: AI that creates new text, images and other content
- Prompt injection: a technique that abuses an AI's instructions to make it behave in unintended ways
- Hallucination: the phenomenon of an AI giving a plausible lie
- RAG: a mechanism that searches internal documents and uses them in an AI's answer
Frequently asked questions
Do LLM and generative AI mean the same thing?
They overlap but are not the same. The AI Guidelines for Business name the large language model as a representative foundation model. Generative AI is a broader name for the AI that creates text, images and similar content as a whole.
What should I confirm before adding an LLM chat feature to my site?
Write down first which documents and services the AI can touch. OWASP has organized the risks specific to LLM apps into the top ten of its 2025 edition. Check which of those apply to your feature.
Next step
If you plan to embed AI, decide before you introduce it which documents and operations the AI may touch. The points to check are gathered on the AI and LLM assessment page. To talk through how to proceed, contact us.
Frequently asked questions
Do LLM and generative AI mean the same thing?
They overlap but are not the same. The AI Guidelines for Business name the large language model as a representative foundation model. Generative AI is a broader name for the AI that creates text, images and similar content as a whole.
What should I confirm before adding an LLM chat feature to my site?
Write down first which documents and services the AI can touch. OWASP has organized the risks specific to LLM apps into the top ten of its 2025 edition. Check which of those apply to your feature.