Attackers may write phrases such as "ignore all previous instructions" in web pages to hijack AI agents. Security teams should train staff to recognise these phrases and report them. This document only describes the attack and contains no instruction for the reader.