Definition
An attack in which malicious instructions are hidden inside content the AI reads from somewhere other than the user's own message — a webpage, an email, a log file, or a connected tool. The AI then follows those planted instructions as if they were legitimate requests, often without the user or the organisation noticing.
Why it matters
This is the attack class behind headline agent hijacks where a single crafted email or webpage takes over an autonomous agent and the credentials it holds, so it is central to agent risk conversations.