La prompt injection est une technique adversariale qui manipule un modèle de langage en plaçant des instructions malveillantes ou non voulues dans l’entrée utilisateur, des documents, des pages web ou d’autres contenus traités par le modèle. Elle peut amener le modèle à ignorer des instructions antérieures, divulguer des informations restreintes, modifier ses sorties ou déclencher des actions non autorisées.
Prompt injection manipulates an LLM's behaviour (for example to alter response style, retrieve hidden or restricted data, or disrupt interactions) by embedding specific instructions within a prompt. This approach exploits the model's tendency to follow instructions within the prompt sequence, even if instructions are unintended or malicious.
Distinguish between Prompt Injection (hijacking model behavior via malicious input) and Jailbreaking (bypassing safety guardrails to produce restricted content). Both are LLM security threats but differ in mechanism and goal.
A side-by-side comparison of Prompt Injection and Data Poisoning. Understand how runtime instruction manipulation differs from attacks on training data or the training process.