Glossary

What is Prompt Injection?

Input given to a large language model to override its original instructions or steer it into unintended actions. It is one of the main security risks of LLM applications.

In a prompt injection attack, the input aims to override the model’s system instructions or expose confidential information. In the direct type, the user writes the malicious instruction themselves; in the indirect type, the model processes hidden instructions that come from an external source such as a web page or a document.

The risk grows as the model gains access to tools and data. There is no single definitive fix; defense is layered: input and output controls, least privilege, limiting the actions the model can take, and human approval at critical steps.

An AI Gateway can apply some of these controls centrally. The risk matters especially in architectures that bring in external content, such as RAG. The topic also appears in OWASP’s list for LLM applications: OWASP Top 10 for LLM Applications.

← All terms

Related solutions

Related articles

Let's plan your critical systems together

Tell us what you need; our team will work out the right architecture and migration path with you.