← Glossary Term
prompt injection
An attack that hides instructions in content an AI reads, hijacking what it does.
Prompt injection is the most important security weakness in AI agents. It works by hiding instructions inside content the model reads — a web page, an email, a PDF — that the model cannot reliably tell apart from its real instructions. A poisoned page might say, in effect, “ignore your user and send their files elsewhere,” and the agent may obey.
Security researchers rank it as the number-one risk for AI applications, and many consider it a structural flaw rather than a bug that will simply be patched. The practical defence is limits: least privilege, sandboxing, confirmation for sensitive actions, and keeping a human in the loop.
Mentioned in