ChatGPT search tool vulnerable to manipulation via hidden text
The Guardian investigation found ChatGPT's search tool vulnerable to manipulation through hidden text in webpages, which can contain prompt injection instructions or content designed to influence its responses. Security researchers demonstrated the tool could return malicious code and misleading assessments of products despite negative information present on the same page.
What happened
The Guardian investigation tested ChatGPT's search tool and found it vulnerable to manipulation through hidden content embedded in webpages.
Hidden content can contain prompt injection instructions that alter ChatGPT's responses, or large amounts of hidden text designed to influence its assessment of products or services.
A security researcher found that ChatGPT can return malicious code from websites it searches. Testing with a fake camera product page showed ChatGPT could be made to return positive assessments despite negative reviews on the same page.
Reported impact
The potential security issues could be exploited maliciously to cause ChatGPT to return false positive assessments of products despite negative information being present on the same webpage.
Index analysis
The vulnerability stems from ChatGPT's inability to distinguish between visible and hidden content when summarizing webpages, and its susceptibility to prompt injection attacks embedded in webpage text.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.