Executive Summary
WithSecure showed prompt injection against Gemini-based Google Drive features could make AI agents leak sensitive documents and metadata; Google implemented fixes.
What Happened
WithSecure researchers showed how prompt injection attacks against Gemini-based Google Drive features could be used to trick AI-powered agents into leaking sensitive documents and metadata from users’ cloud storage.[rich_content:1] The attack chains abused the model’s tool usage and instructions embedded in documents to bypass user intent and exfiltrate data without traditional malware.[rich_content:1] Google acknowledged the findings and implemented mitigations while emphasizing the importance of defense-in-depth and user controls around AI-powered workspace tools.[rich_content:1]
Affected Products
- Google Gemini
- Google Drive
Vulnerability / Attack Technique
Prompt injection via documents leading to data leakage
Technical Mechanism
Instructions embedded in documents redirected the AI feature's tool use.
AI Security Category
LLM01:2025 Prompt Injection
Part of the Prompt Injection topic.
Relevant Frameworks
RealGround Analysis
Shared documents can carry instructions to the assistant. RealGround analysis: be careful with AI features on folders containing externally shared files.
Recommended Controls
- Apply Google's fixes
- Separate externally shared files from sensitive folders
Related RealGround Service
Secure AI Agent Build · or start with the free AI Agent Security Scanner.
Sources
Summary and analysis written by RealGround from the source above; last updated 2026-10-18.