# Prompt Injection

An attack technique where malicious instructions are hidden in input given to an AI system to make it behave against its intended purpose.

**Prompt injection[↗](/prompt-injection)** is an attack technique where malicious instructions are hidden inside input given to an AI system, often disguised within a document, webpage, or message the AI is asked to process, in order to make it ignore its original instructions and behave in a way the attacker wants instead. It's conceptually similar to older injection attacks against traditional software, but targets the AI's language-based instructions rather than code.

A real-world example is a hidden instruction embedded in a webpage that an AI assistant is asked to summarize, which secretly tells the AI to leak private data or take an unintended action instead of doing the summary. As AI systems are increasingly connected to email, files, and other tools, defending against prompt injection, through input filtering, strict permission boundaries, and careful system design, has become one of the most active areas of AI security research.