Prompt injection
Prompt injection is an attempt to make a language-model application treat attacker-controlled input as instructions, redirecting its behavior away from the intended task or rules.
Source material is not permission
An assistant may read a document to answer your question. That document can also contain text trying to tell the assistant what to do next. The problem starts when the application treats those instructions as authority rather than as material to examine.
OWASP distinguishes direct injection, supplied through prompt input, from indirect injection in external material such as a webpage or file. Microsoft also describes instructions arriving in tool results. Ordinary user instructions are not attacks simply because they change an answer; the issue is whether the input overrides the application’s intended task or rules.
The impact depends on what the application can do
An injection might distort an answer or influence which source the assistant recommends. An application with connected tools may also be pushed toward an unintended action. Reading an injected document does not, by itself, prove that a file was sent or an account changed.
Treat model output as something the application must check before acting on it. Limit tool permissions, separate external content from instructions, and require appropriate authorization for sensitive actions. OWASP recommends several layers of protection; a warning in a prompt alone is not a complete defense.
For someone reviewing an assistant’s proposed action, ask where it came from: your request, or the document being read? Check the destination and information involved before approving it. The AI document article follows that decision and distinguishes a proposed action from a completed transfer.