retrieval poisoning

Retrieval poisoning is a security vulnerability where short, manipulated passages in user-generated web content can mislead deep-research AI agents. A May 2026 preprint from Cornell Tech researchers Tingwei Zhang, Harold Triedman, and Vitaly Shmatikov demonstrated that as few as 13 words in a Reddit-style comment can steer an AI agent's output. This technique exploits the retrieval step in AI search, making the model only as trustworthy as the pages it retrieves. The core issue is that AI search has recreated an old web-security problem: the easiest pages to manipulate are often the ones retrieved. This tag covers discussions of retrieval poisoning in the context of AI security, web content manipulation, and the integrity of AI-driven research tools on Windows platforms.