OpenAI's ChatGPT search tool, an AI-powered browsing feature launched this month, is vulnerable to generating misleading summaries, according to a recent report by The Guardian.

The tool, designed to streamline browsing by summarising web pages, was found to be susceptible to manipulation. The Guardian demonstrated that by embedding hidden content in websites it created, ChatGPT Search could be tricked into producing "entirely positive" summaries, even when the original content included negative reviews. This method also known as "prompt injection" can be done with hidden contents having instructions from third parties which can alter ChatGPT's responses. The method also enabled the generation of malicious code.

Hidden content attacks are a recognised vulnerability in large language models. The report highlighted that Google, a dominant player in the search market, has greater experience addressing similar challenges.