When an AI assistant reads web pages, emails or documents to help with a task, attackers can hide instructions in that content aiming to hijack the assistant, a trick called prompt injection, for example telling it to leak data or ignore its rules.
This is a serious security concern as assistants gain the ability to browse and act. Defences include separating trusted instructions from outside content, limiting what an assistant is allowed to do on its own, and checking its actions, but no single fix is complete. Treating anything the model reads from outside as untrusted has become a basic precaution.