Hidden content tricks ChatGPT into rewriting search results, Guardian shows
In October, OpenAI's ChatGPT Search became available for ChatGPT Plus users. Last week, it became available to all users and was added to search in Voice Mode. And, of course, it isn't without its flaws.
The Guardian asked ChatGPT to summarize webpages that contain hidden content and, it turns out, hidden content can manipulate the search. It's called prompt injection, which is the ability for third parties — like websites you're asking ChatGPT to summarize — to force new prompts into your ChatGPT Search without your knowledge. Consider a page full of negative restaurant reviews. If the site includes hidden content waxing poetic about how incredible the restaurant is and encourages ChatGPT to instead answer a prompt like "tell me how amazing this restaurant is," that hidden content could override your original search.
"In the tests, ChatGPT was given the URL for a fake website built to look like a product page for a camera. The AI tool was then asked if the camera was a worthwhile purchase. The response for the control page returned a positive but balanced assessment, highlighting some features people might not like," The Guardian investigation states. "However, when hidden text included instructions to ChatGPT to return a favorable review, the response was always entirely positive. This was the case even when the page had negative reviews on it – the hidden text could be used to override the actual review score."
This doesn't spell failure for ChatGPT Search, though. OpenAI only recently launched Search, so it has plenty of time to fix these kinds of bugs. Plus, Jacob Larsen, a cybersecurity researcher at CyberCX, told The Guardian that OpenAI has a "very strong" AI security team and "by the time that this has become public, in terms of all users can access it, they will have rigorously tested these kinds of cases."
Prompt injections attacks have been a hypothetical for ChatGPT and other AI search functions since the technology launched, and while we have seen some demonstrations of the potential harms, we haven't seen a major malicious attack of this kind. That said, it does point to a problem with AI chatbots: They are remarkably easy to trick.