ChatGPT for Google Sheets Exposes Workbook Data Leak Vulnerability

A serious security vulnerability has been discovered in ChatGPT for Google Sheets, an extension released by OpenAI. A single Indirect Prompt Injection attack occurring in just one sheet can affect a user's entire account.

This attack exploits permissions granted by the user, and can cause the following damages simultaneously:

• Data leakage from multiple workbooks within the account
• Display of interactive phishing popups
• Replacement of the GPT sidebar with an interface controlled by the attacker
• Unauthorized modification of workbooks

The attack mechanism works by having untrusted data contained in external data sources (such as imported sheets or ChatGPT connectors) manipulate ChatGPT, causing it to execute an attacker-controlled external script using permissions granted by the user.

This vulnerability has been reported to OpenAI, but no substantive response or documentation update has been made to date. Existing documentation only focuses on the model's functional limitations or data handling methods, and does not sufficiently explain security risks arising from model manipulation.