Gemini hack found with GeminiJack

Gemini hacking has been found with GeminiJack, allowing hackers to access data on Gmail and Docs.

AI (Artificial Intelligence or Artificial Intelligence) is so useful that it has caught the eye of the whole world. Many companies have invested in its development, but many of the models available today have detected many weaknesses, many of which can harm users because they can be used by hackers.

According to a report by the website Cyber Security News, a research team from Noma Labs, an AI-based risk specialist firm, discovered a way to hack Google's Gemini AI tool on the Enterprise version (a version for large business enterprises) through a functional security vulnerability without relying on victim interaction, or Zero-Click, which the research team named the vulnerability "GeminiJack," and described the vulnerability as an architectural flaw of the system rather than a Vulnerability, according to the general understanding that the vulnerability works in terms of Prompt Injection: After the victim opens a Google Doc file, an email, or an invitation to save an appointment to a calendar (Calendar Invite) where Prompt is secretly hidden, which leads to Exfiltration of data saved on Google Workspace that the hacker wants to hacker.

The research team went on to explain that the problem is in the architecture of Google Gemini Enterprise's RAG (Retrieval-Augmented Generation) system, which is indexing, the Querries petition in the Email, Calendar events, and documents on Google Docs, to be quick to find and manage. As a result, hackers can poison AI to find sensitive information through "confidential," "API key," and "acquired" search terms on the Workspace platform. What is searched for will be displayed in Image format. Tag in the HTML language and then export it to hackers with the HTTP protocol, where the steps to perform it will contain the following 4 steps.

1. Poisoning Hackers will share documents in Docs, Email, or Calendar invitations with Prompt embedding with keywords like "Search" Sales, and the Img Tag will be displayed.

2. Trigger, because the AI Gemini has been poisoned from the previous stage. If the employee in the victim organization performs a search with a normal command like "Sales docs?" the first step will work immediately.

3.Retrieving the RAG mechanism of the Gemini body

4.Send to the hacker (Exit) The AI will retrieve everything relevant from every data source, convert it into a Tag Image to the hacker has stolen it to use.

This method is called very difficult to detect, because for employees, sending such a command is like a common command that nothing is wrong. And in the corner of the security system, it is not counted as a phishing scam. There is no malware release on the machine.

The source did not indicate what Google's opinion on the issue or whether Google has resolved the issue, so the organization running Google Gemini Enterprise may face the risk of unwitting and difficult data theft for at least a while.

# Recap 2025 # Take care of yourself # Open statements # Includes IT matters # googlegemini