ChatGPT search found vulnerable to manipulation

The AI search tool is misled by hidden website text.

Open AI, ChatGPT

New research by The Guardian reveals that ChatGPT Search, OpenAI’s recently launched AI-powered search tool, can be misled into generating false or overly positive summaries. By embedding hidden text in web pages, researchers demonstrated that the AI could ignore negative reviews or even produce malicious code.

The feature, designed to streamline browsing by summarising content such as product reviews, is susceptible to hidden text attacks—a well-known vulnerability in large language models. While this issue has been studied before, this marks the first time such manipulation has been proven on a live AI search tool.

OpenAI did not comment on this specific case but stated it employs measures to block malicious websites and is working to improve its defences. Experts note that competitors like Google, with more experience in search technology, have developed stronger safeguards against similar threats.