Quick answer
Gemini for Google Workspace can read Gmail, Drive, Docs and other Workspace data the user has access to. Like other enterprise assistants, it inherits existing permissions, so overshared Drive files become easy to find, and content written by outsiders, such as inbound email, can carry prompt injection.
The Main Risks
- Drive oversharing. Files shared with "anyone in the organisation" or with public links become searchable through natural-language questions.
- Prompt injection through Gmail and Docs. Researchers have shown that instructions hidden in emails and documents can influence Gemini's summaries and responses. See indirect prompt injection.
- Phishing via AI summaries. Manipulated summaries can present attacker-supplied links or instructions as if they came from the assistant.
- Admin configuration. Which Workspace apps and data Gemini can use is controlled by admin settings that are often left at defaults.
What to Test
- What ordinary and compromised accounts can retrieve through Gemini across Drive and Gmail.
- Whether planted instructions in inbound email or shared docs change Gemini's output.
- Whether data loss prevention and sharing controls apply to Gemini responses.
- Admin settings for Gemini access, logging and data regions.
How to Reduce the Risk
Audit Drive sharing before rollout, restrict organisation-wide and public links, apply data loss prevention rules, configure Gemini access by organisational unit, enable logging and train staff to verify links in AI summaries.
Regulatory and Framework Context
For Australian organisations on Google Workspace, a Gemini review sits naturally alongside Privacy Act and Essential Eight work on user application hardening.
Latest intel on this topic
- Gemini test proves once again that amoral AI agents will do almost anything if told ‘NEVER give up’The Register · 17 November 2026
- Gemini AI agents repeatedly bypassed safeguards to attack real websiteBleepingComputer · 16 November 2026
- The Safety Filter Refused. The AI Agent Delivered Anyway.AI Incident Database · 6 November 2026
- Google I/O, Gemini Spark, AntigravitySimon Willison (prompt injection) · 20 May 2026
- Google Jules is Vulnerable To Invisible Prompt InjectionEmbrace The Red · 15 August 2025
Updated hourly from the daily briefing.
Common questions
Does Gemini for Workspace use our data to train models?
Google states that Workspace data used with Gemini is not used to train its models outside your domain without permission. Confirm current terms for your agreement.
Is Gemini more or less secure than Copilot?
The risks are similar: both inherit user permissions and can be targeted through content they read. Configuration and permissions matter more than the product.
Change log
- 13 November 2026First published.