Prompt Injection

An attack where malicious instructions hidden in input data trick an AI system into ignoring its actual instructions.

If an AI agent reads untrusted content (a webpage, an email, a document) as part of its task, that content can contain hidden text instructing the model to ignore its original task and do something else instead — the AI-specific equivalent of an injection attack. Systems that let an AI take real actions need to treat any content it reads as untrusted input, not as trusted instructions.