ChatGPT Search, OpenAI’s recently launched AI-powered search engine, has been found susceptible to hidden text attacks, which can lead to misleading summaries of webpage content, according to an investigation by The Guardian.
The feature, designed to enhance browsing by providing summaries of webpage content, including product reviews, has been shown to be tricked by hidden text embedded in websites. This manipulation can cause the AI to produce overly positive summaries by ignoring negative feedback or, in some cases, generate malicious code when prompted.
Hidden text attacks exploit vulnerabilities within large language models, such as the one used in ChatGPT, by feeding the system with manipulated input. While these risks have been well-documented in AI research, The Guardian suggests that this incident may be the first instance of such manipulation being used on a live AI search engine.
In comparison, Google, a search engine leader, has a much longer history of tackling similar issues and is considered better equipped to handle these challenges, given its experience and the robustness of its algorithms.
OpenAI, the company behind ChatGPT Search, declined to comment on the specific incident but stated that it employs various strategies to block malicious websites and continues to refine its safeguards to improve the platform’s reliability.
This discovery highlights growing concerns over the reliability of AI-driven search products, particularly as they become more integrated into online experiences. As these tools are increasingly relied upon, ensuring they are resilient against manipulation and misinformation is essential for maintaining their integrity and trustworthiness in delivering accurate information.
