Google Cloud Platform (GCP) Gemini Cloud Assist Prompt Injection Vulnerability

Tenable Research discovered that Gemini Cloud Assist was susceptible to prompt injection via log content. When the user analyzed these logs with Gemini Cloud Assist, the attacker could potentially exploit Gemini integrations or trick the user into visiting a legitimate-looking phishing link.

This prompt injection finding enabled attackers to embed malicious prompts within a victim’s logs. When a user analyzed these logs using the Gemini Cloud Assist feature, the malicious prompt was processed, leading to one or more potential outcomes:

• Exploit of Gemini Integrations: Attackers may execute unauthorized actions or gain elevated access by leveraging integrations tied to the Gemini system.
• Phishing Risks: The malicious prompt could trick users into visiting legitimate-looking but fraudulent phishing links, thereby compromising sensitive information.

The attack remained covert because the injected payload was hidden in the log details. The victim must click “Additional prompt details” to view the full log entry, including the malicious content.

The finding could be abused by a victim that sends a prompt that triggers Gemini to look for logs in any way. An example of such a prompt would be: "What was the latest cloud function execution in my environment?"

Steps to Reproduce:

1. An attacker injects a crafted log entry containing a malicious prompt through user input control such as a User-Agent HTTP header.

2. The victim reviews logs via the GCP Log Explorer using Gemini Cloud Assist.

3. The summarization feature processes the injected prompt, executing and rendering the attacker’s payload.

4. The user is inadvertently misled into performing actions beneficial to the attacker (e.g., clicking phishing links or enabling unauthorized access).