British media found that ChatGPT's search has vulnerabilities, and hidden text can affect response content

ChainCatcher news reports that a investigation by the UK's The Guardian has found potential security issues with OpenAI's ChatGPT search tool, which could be exploited to manipulate content by hiding it.

The Guardian tested ChatGPT's response when asked to summarize web pages containing hidden content. The hidden content may include third-party instructions, such as extensively discussing the benefits of a particular product, which can alter ChatGPT's responses. This technique can be maliciously used, for instance, to make ChatGPT return positive reviews despite negative comments about the product on the same page. Testers provided ChatGPT with a fake URL that resembled a camera product page and then asked if the camera was worth buying. ChatGPT gave a positive but balanced review, while highlighting some features that people might not like. However, when the hidden text contained instructions for ChatGPT to return positive reviews, its responses were always entirely positive, even when there were negative comments on the page.

CyberCX cybersecurity researcher Jacob Larsen stated that if ChatGPT search remains open to everyone in its current state, there could be "high-risk" individuals creating websites specifically designed to deceive users. (Jin Shi)