Key terms
What is prompt injection?
Prompt injection is an attack in which text in the input, such as a web page or document, contains instructions that manipulate an AI system into ignoring its intended instructions or taking unintended actions. It is a particular risk for agents that read untrusted content.
The full answer
Prompt injection is an attack in which text in the input, such as a web page or document, contains instructions that manipulate an AI system into ignoring its intended instructions or taking unintended actions. It is a particular risk for agents that read untrusted content.
The usual defenses are narrow permissions, short-lived access and an approval gate before consequential actions.