What feeds an AI agent reads shapes what it decides?

LLM agents consume ranked feeds (search results, social posts, email) before deciding, yet safety tests ignore this layer entirely. Researchers held model, prompt, and persona constant while varying only feed composition across 2,785 decision rollouts on four models. They found adversarial feeds can completely overturn uncertain decisions (up to 95-point swings), yet cannot dislodge firmly held positions—a one-way asymmetry. The effect survived multiple controls and generalized to security-critical choices like removing deployment gates.