Skip to content
Submit incident
AII-1001-37437
Well documentedThree or more independent publishers report the core facts.

Google Gemini vulnerabilities allowed data theft via prompt injection

Tenable Holdings disclosed three indirect prompt injection vulnerabilities in Google Gemini tools that could allow attackers to steal user data and trigger unauthorized actions. The vulnerabilities affected Gemini Cloud Assist, Gemini Search Personalization Model, and Gemini Browsing Tool, and Google has since patched them.

30 September 2025 Β· Google LLC Β· Security and cyberattacks
30 September 2025
AII-1001-374376 sources ↓
Quick facts
Category
Security and cyberattacks
Country
United States
Source
Cyber Security News
Date
30 September 2025
Drafted by
Raidu Incident Detection

What happened

Tenable Holdings discovered three vulnerabilities in Google Gemini tools dubbed the 'Gemini Trifecta,' all involving indirect prompt injection attacks. Gemini Cloud Assist was vulnerable to attackers poisoning log data by inserting malicious payloads such as manipulated User-Agent headers into Cloud Logging, which would then execute when Gemini summarized the logs.

The Gemini Search Personalization Model could be exploited by injecting malicious queries into a victim's Chrome search history via JavaScript from a malicious website, which Gemini would then treat as legitimate context. The Gemini Browsing Tool could be manipulated to make outbound requests to attacker-controlled URLs by crafting prompts in Gemini's internal language, silently exfiltrating sensitive data to the attacker's server.

Google has since addressed these vulnerabilities.

Reported impact

Attackers could generate phishing links within log summaries, query sensitive cloud assets, trigger unauthorized actions, or manipulate Gemini's behavior to extract sensitive data such as saved personal information or location details. Data exfiltration via the Browsing Tool could occur undetectably to users, as the sensitive data would be captured silently through background tool execution.

Index analysis

The vulnerabilities highlight risks of indirect prompt injection, where malicious instructions are hidden in trusted data sources that the AI system later processes. The attacks underscore how AI integrations can become active threat surfaces if not properly secured, requiring security practices commensurate with traditional enterprise infrastructure.

Classification

Category
Security and cyberattacks
Industry
Digital security
Country
United States
Harm status
Potential harm
Intent
Unintentional
Timing
Post-deployment
Harmed
Consumers, Business
Failure modes
Prompt injection, Data leak
AI systems
Gemini Cloud Assist, Gemini Search Personalization Model, Gemini Browsing Tool

Sources and evidence

This record was researched and written by the Index, based on the following original reporting.

Cyber Security News
Google Gemini Vulnerabilities Let Attackers Exfiltrate User's Saved Data and Location
30 September 2025
SiliconANGLE
'Gemini Trifecta' vulnerabilities in Google AI highlight risks of indirect prompt injection - SiliconANGLE
30 September 2025
Dark Reading
'Trifecta' of Gemini Flaws Turn AI into Attack Vehicle
30 September 2025
Global Security Mag Online
Tenable exposes AI flaws in Google Gemini that could have let hackers steal data from millions - Global Security Mag Online
30 September 2025
Infosecurity Magazine
Gemini Trifecta Highlights Dangers of Indirect Prompt Injection
30 September 2025
Cybernews
Hackers find hidden exploit in Google's Gemini AI | Cybernews
1 October 2025