Safety
Prompt injection
An attack where instructions hidden in content the model reads, such as a web page, email or document, try to hijack its behaviour.
Prompt injection is a key risk for agents that browse or read untrusted input. Treat external content as data, limit what tools an agent can use, and require confirmation for sensitive actions.
Frequently asked questions
What does Prompt injection mean?
An attack where instructions hidden in content the model reads, such as a web page, email or document, try to hijack its behaviour. Prompt injection is a key risk for agents that browse or read untrusted input. Treat external content as data, limit what tools an agent can use, and require confirmation for sensitive actions.