Hidden ads in search results steered what a shopping AI bought

Researchers showed that hidden instructions inside web pages can redirect an AI shopping agent's choices without the agent or the human noticing. The attack works by placing text inside product listings or search results that the agent reads as part of the page. The agent then steers toward the advertised item, even when better options exist. Tests on real product data confirmed the method works at scale. The deeper problem is that humans watching the agent have no way to see this happening. The agent's reasoning looks normal from the outside. Nothing in the interface flags that a third-party seller wrote instructions the agent followed. The human approved a task at the start, walked away, and received an outcome shaped by an unseen party. For teams building shopping or research agents, this exposes a gap in what humans can inspect after an agent finishes a job. Knowing what the agent did is not enough if you cannot see what the agent read and what weight it gave each source. Without that, humans reviewing the agent's work cannot tell a good recommendation from a steered one.