Prompt injection

Text inside a document, a web page or a tool result that the model reads as an instruction rather than as content.

how it works · the vocabulary

The model cannot tell your instructions from the ones an attacker left in a file it was asked to summarise. It is the defining security problem of agents, because an agent with tools can act on what it reads. Treat everything a model retrieves as untrusted, and never let a tool result authorise an action.

Nearby