Prompt Injection
The behavior of the model by hiding the malicious instructions into the legitimate input.
Prompt injection is particularly risk in AI agents that read external data source (web, document); secret instructions may cause deviation from the main task of the model.
Member comments