AegisAI Perspectives · 165
A startup protects Gemini in Gmail from prompt injection in email with Google's defenses, Admin console limits, and a check on each email for hidden text.
See it on your own Google Workspace or Microsoft 365 tenant. No MX change.
Or get the Q3 2026 State of the AI Attack report (free, instant download)
Not ready for a demo? Check your domain.

A startup protects Gemini in Gmail with Google's built-in defenses1, Admin console limits on where Gemini reads mail5, and a check on each incoming email for hidden instructions21. AI assistants such as Gemini read mail for the user, and Google says indirect prompt injection hides malicious instructions in external data such as emails3. AegisAI inspects incoming mail for the concealed payloads engineered to manipulate them21.
If you run IT at a startup and want to protect Gemini in Gmail from prompt injection in email, start with the Google settings below. The email check and the eight general steps for any AI email assistant are in how to protect AI email assistants from prompt injection. How the Google Workspace connection works is on the inbound email security page. Background on Gemini and Copilot as mail readers is in The Inbox Has a Second Reader.
Key takeaways
- Google lists six layers of defense for Gemini against indirect prompt injection, from content classifiers to model resilience.
- Gemini features in Workspace services, Gmail included, are on by default4.
- Hidden text set to font-size:0 or color:white shows as blank space in Gmail, yet Gemini's summary can repeat it as a fake security alert18.
- The UK NCSC says to reduce the risk and the impact of prompt injection, and to beware any vendor that claims it can stop it17.
| Question | Answer | Source |
|---|---|---|
| How does a startup protect Gemini in Gmail from prompt injection in email? | Google's built-in defenses, Admin console limits, and a check on each incoming email. | Google Workspace Admin Help; AegisAI |
| What does Google build into Gemini against prompt injection? | Six layers, from prompt injection content classifiers to model resilience. | Google Workspace Admin Help |
| Where is Gemini turned off for Gmail? | Menu, Generative AI, Gemini for Workspace, then the Feature access panel, Edit next to the service, On or Off, Save. | Google Workspace Admin Help |
| Does turning Gemini off in Gmail keep Gemini out of mail? | Not fully7. See step 4 and the questions below. | Google Workspace Admin Help |
Short on time? Watch the explainer, or book a demo.
It reaches Gemini through the email body. Google's security blog describes indirect prompt injection as hidden malicious instructions within external data sources, which may include emails, documents, or calendar invites3. OWASP says indirect prompt injections occur when a model accepts input from external sources, such as websites or files15.
In a write-up dated July 10, 2025, 0DIN showed instructions styled with font-size:0 or color:white. Gmail rendered them as blank space. Gemini's summary repeated them as a fake security alert.
The root cause is in the model. The UK NCSC says there is no inherent distinction between data and instruction in a large language model16. So an email that Gemini reads is also text that Gemini may follow. Treat every summary of external mail as untrusted output.
Google runs six layers inside Gemini. They are prompt injection content classifiers (machine-learning models that detect malicious prompts and instructions), security thought reinforcement (security instructions added around the prompt content), markdown sanitization and suspicious URL redaction (redacting external image URLs and suspicious links), a user confirmation framework, end-user security mitigation notifications, and model resilience (which protects Gemini models from explicit malicious manipulation)1.
Two gaps stay open for a startup.
The NCSC advice is to reduce the risk and the impact. Plan for fewer injected emails reaching Gemini and for less that Gemini can reach when one does.
Start with the edition, then set Gemini in Gmail access, admin rights, the Gemini app, smart features, mail flow protections and user guidance. Do them for one organizational unit first, then roll out.
Confirm the edition. Google lists Enterprise Standard and Enterprise Plus, among others, as the editions that can manage Gemini feature access per Workspace service14. Check yours before you plan per-team rules.
Decide who needs Gemini in Gmail. Go to Menu, Generative AI, Gemini for Workspace, open the Feature access panel, click Edit next to Gmail, choose On or Off, then Save.
Give the change to the right admin. It needs the Gemini Settings administrator privilege, and group settings override organizational units6. Check group membership before you test.
Set the Gemini app's reach into Gmail. Go to Menu, Generative AI, Gemini app, then Apps. Turn Workspace apps off for any organizational unit where the Gemini app should not read Gmail12. It is on by default and lets the Gemini app find information in Gmail. Changes can take up to 24 hours13.
Check smart features. Users need smart features and personalization turned on to use Gemini in Workspace services8. A super administrator sets the default under Menu, Account, Account settings, Smart features for Google Workspace9.
Turn on Gmail's mail flow protections. Under Apps, Google Workspace, Gmail, Safety, choose an action for each attachment and spoofing option. The default is Keep email in inbox and show warning, and the other choices are Move email to spam and Quarantine10. Turn on enhanced pre-delivery message scanning under Gmail, Spam, Phishing and Malware11.
Tell users what an injected summary looks like. A Gemini summary that warns of a security problem can repeat hidden text from the email itself. Ask users to forward that email to IT instead of acting on it.
Write the date and the organizational unit next to each step as you finish it.
Set each control on the screen in the second column, and compare it with the default in the third. "Not stated" means the Google page gives no default.
| Control | Where in the Admin console | Default | What it changes for prompt injection |
|---|---|---|---|
| Gemini in Gmail | Menu, Generative AI, Gemini for Workspace, Feature access | On | Removes Gemini from Gmail for that unit or group |
| Gemini app access to Workspace apps | Menu, Generative AI, Gemini app, Apps | On | Stops the Gemini app finding information in Gmail |
| Smart features for Google Workspace | Menu, Account, Account settings | Off for domains based in the European Economic Area, Japan, Switzerland or the UK, on in all other regions; a super administrator can set it | Gemini in Workspace services needs it on |
| Gmail Safety protections | Apps, Google Workspace, Gmail, Safety | Keep email in inbox and show warning, for each attachment and spoofing option | Checks attachments and spoofing at mail flow |
| Enhanced pre-delivery message scanning | Apps, Google Workspace, Gmail, Spam, Phishing and Malware | Not stated | Delays suspicious mail for extra checks before delivery |
| Google's six prompt injection layers | Inside Gemini | Built in | Built into Gemini by Google |
Screenshot each screen after you set it, and keep the screenshots with your access review.
Tested against the real thing, reproducing the disclosed 0DIN payload in ordinary benign mail, AegisAI flags it and identifies the hidden text as an injected instruction rather than matching a keyword23. Hidden regions of the body stay in scope, so an instruction written for Gemini or Copilot is read and judged22. Reads mail through the Microsoft 365 or Google Workspace API after delivery25. Runs in monitoring mode first. Removal needs write access your admin grants and can revoke.
Measure the question "How does a startup protect Gemini in Gmail from prompt injection in email?" with these three records, taken on day 1.
To see which hidden instructions reach your own Gmail, use the demo request page. Compare the three on day 30.
This guide groups the controls into three layers. They are Google's six built-in defenses, Admin console limits on where Gemini reads mail, and a check on each incoming email for hidden instructions. AegisAI inspects incoming mail for the concealed payloads engineered to manipulate them. The UK NCSC says to reduce the risk and the impact of prompt injection.
Not on the Google pages cited here. The Admin console controls on those pages turn Gemini on or off per service and set the Gemini app's reach into Gmail. The six layers, from prompt injection content classifiers to model resilience, are built into Gemini.
Yes. Google says the default setting for Gemini features in Workspace services is on. Turn it off per unit or group under Generative AI, Gemini for Workspace, Feature access.
Not fully. Google says users can still access an app's data when using Gemini in other apps.
In the 0DIN case, the summary showed a fake security alert built from text the reader could not see. Treat any urgent alert in a summary as coming from the sender.