ChatGPT search tool could be open to manipulation using hidden content: Report

ChatGPT search tool could be open to manipulation using hidden content: Report

The search tool of OpenAI’s ChatGPT might be open to manipulation through the use of hidden talent, besides returning malicious codes from across the websites that it searches, a new investigation has revealed.

The research, conducted by The Guardian, tested the AI chatbot’s response to a query about summarising the web pages containing hidden content.

During the testing, it was found that the hidden content could contain instructions from third parties which can alter its responses. This is also called ‘prompt injection’. Also, it may contain content that has been specifically designed to influence the response of the AI chatbot.

It must be noted that ChatGPT has been made available to paying customers, while OpenAI has urged users to consider making it their default search tool.

What probe suggests

Through its research, The Guardian noted that techniques, like prompt injection, can be used maliciously by some people. Giving an example, it stated that this can cause ChatGPT to provide a positive assessment of some product, which otherwise has negative reviews on the same page.

“A security researcher has also found that ChatGPT can return malicious code from websites it searches,” read the report.

Key takeaways

During the research, the ChatGPT was provided with a fake website’s URL specifically designed to look like a product page for a camera. After being asked whether to buy the camera or not, the AI chatbot in reply gave a “positive but balanced assessment, highlighting some features people might not like,” the report noted.

The response, however, came out only positive when hidden text included instructions to it to return a favourable review. This was noted even when the page had negative reviews for the product, highlighting how hidden text might get used to “override the actual review score.”

If the ChatGPT search system is released fully in its current state, there is ‘high risk’ of people coming up with websites that are specifically aimed towards deceiving users, said Jacob Larsen, a cybersecurity researcher at CyberCX.

Larsen added, “This search functionality has come out (recently) and it’s only available to premium users… “They’ve got a very strong (AI security) team there, 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”.

Nitish Reddy Overheard Relatives Cursing His Dad, Friends Duped Money – How IndiaGot A Star Previous post Nitish Reddy Overheard Relatives Cursing His Dad, Friends Duped Money – How IndiaGot A Star
2024 box office hits and misses: ‘Wicked’ soars, ‘Madame Web’ flops Next post 2024 box office hits and misses: ‘Wicked’ soars, ‘Madame Web’ flops

Leave a Reply

Your email address will not be published. Required fields are marked *