ChatGPT Search, a new AI-powered search engine launched this month, has been exposed by U.K. newspaper The Guardian for producing misleading summaries.

The search feature of ChatGPT is designed to help users browse faster by summarizing product reviews on web pages. However, The Guardian discovered that by adding hidden text to websites, it could manipulate ChatGPT into disregarding negative reviews and generating “completely positive” summaries. Additionally, ChatGPT Search could be tricked into producing malicious code through this method.

While hidden text attacks are a known risk for large language models, this is the first instance of it being demonstrated on a live AI-powered search product. Google, the leading search engine, has more experience in handling similar issues, as pointed out by The Guardian.

OpenAI did not comment on this specific incident when contacted by TechCrunch, but mentioned that it employs various methods to prevent malicious websites and is constantly enhancing its security measures.