A new preprint on arXiv warns that as people hand off option-ranking tasks—like shopping, job hunting, or travel planning—to AI agents, those agents become targets for manipulation. The authors introduce the concept of "agent ads": content crafted specifically to steer an agent's ranking output, rather than to persuade a human directly.

The paper, titled "Welcome to the Age of Agent Ads," focuses on "on-policy attacks" as a way to exploit the agent's decision-making process. The abstract notes that defenses against such steering attacks are expected, but the full details of the proposed defense are not included in the available summary.

Because only the abstract is available, this write-up reflects that limited scope. The authors do not compare their approach to other attack or defense methods in the provided text, so no such comparison is made here.