"Agent Goal Hijack occurs when an attacker manipulates an agent's objectives, task selection, or decision pathways through techniques including prompt-based manipulation, deceptive tool outputs, malicious artifacts, forged agent-to-agent messages, or poisoned external data."
An agent doesn't just execute one command. It plans. It decides which tools to use, in what order, toward what goal. That planning process lives in natural language — the same medium attackers can write in.
Hijacking the objective means changing what the agent is trying to achieve. Hijacking task selection means changing which actions it picks. Hijacking decision pathways means changing how it reasons about what to do next. Any of these is enough.
Prompt-based manipulation — instructions injected directly into what the agent reads.
Deceptive tool outputs — a tool the agent calls returns a result that contains hidden instructions.
Malicious artifacts — a document, file, or image the agent processes contains an embedded attack.
Forged agent-to-agent messages — in multi-agent systems, one agent sends fake instructions to another.
Poisoned external data — a webpage, database, or API response the agent retrieves has been tampered with.