AI Security

Protecting Customer Documents in an AI Summarization Workflow with Make AI Agents

Published 2026-08-14 by AISecAll Editorial

TL;DR: Store original files in a private Google Drive folder, grant the Make AI Agent a scoped read‑only role, copy files to an encrypted temporary bucket before processing, run the LLM in a sandboxed Make scenario, and delete the temporary copy immediately after summarization. Log every file access and keep audit records for at least 30 days.

What is the threat model for AI‑driven document summarization?

When an AI agent reads a confidential document, three attack surfaces appear:

• Data at rest: the original file on Google Drive could be accessed by an unauthorized user.
• Data in transit: the file is transferred to the LLM service; interception could expose content.
• Data in the AI service: the LLM may retain excerpts in its context window or logs.

Small teams often overlook the first two because they assume Google Drive’s built‑in permissions are sufficient. The OWASP GenAI Security Project warns that LLM providers can inadvertently cache prompts, making it essential to limit what the model sees and for how long.

How to isolate the AI agent from raw files?

Instead of letting the Make scenario read the original document directly, copy it to a short‑lived, encrypted storage location. The pattern looks like this:

1. Create a temp‑bucket in a cloud provider that supports server‑side encryption (e.g., Cloudflare R2 or AWS S3).
2. Use a Make module to copy the file from Google Drive to the bucket. The copy operation should be performed over HTTPS and the bucket must enforce encryption-at-rest.
3. Pass the temporary object URL to the LLM module. The URL should be a signed, expiring link (e.g., GET valid for 5 minutes).
4. After the LLM returns the summary, delete the temporary object immediately.

This approach guarantees that the LLM never sees the original file path and that the copy disappears as soon as it is no longer needed.

Secure storage and transmission of documents

Two technical controls are critical:

• Encryption in transit: All Make modules communicate over TLS 1.2+. Verify that the signed URL uses https:// and that the LLM endpoint enforces TLS.
• Encryption at rest: Enable server‑side encryption on the temporary bucket and, if possible, client‑side encryption before upload. The key should be managed in a secret‑management service, not hard‑coded in the scenario.

If your organization already uses a key‑management service (KMS), reference the key ID in the Make module’s configuration instead of a raw secret.

Least‑privilege access to Google Drive

Make AI Agents authenticate to Google Drive via an OAuth service account. Follow the principle of least privilege:

• Grant the service account the drive.file scope only, which allows access to files the account creates or explicitly shares.
• Share a dedicated “Input” folder with the service account and place all documents to be summarized there.
• Do not grant drive.readonly or broader scopes; they would expose the entire Drive hierarchy.

Google’s documentation on OAuth scopes is the authoritative source for the exact scope strings.

Audit logging and data retention

Every step that touches a document should be logged in a tamper‑evident store. A simple approach for small teams is to write JSON log entries to a Google Sheet or a Cloudflare Logpush endpoint. Include:

• Timestamp (ISO 8601)
• Document identifier (file ID, not the name)
• Action performed (copy, summarize, delete)
• Actor (service account email)
• Result (success/failure, error code)

Retain logs for at least 30 days to satisfy most data‑protection regulations and to enable forensic analysis if a breach is suspected.

Putting it together: a step‑by‑step checklist

1. Prepare the Google Drive folder: Create AI‑Summaries‑Input and share it with the service account using the drive.file scope.
2. Configure the temporary bucket: Enable server‑side encryption and set a lifecycle rule to delete objects after 1 hour (as a safety net).
3. Build the Make scenario:
   • Trigger: new file in the shared folder.
   • Action 1: copy file to the encrypted bucket, generate a signed URL.
   • Action 2: invoke the LLM via the Make AI Agents module, passing the signed URL as input.
   • Action 3: write the summary back to a separate “Output” folder in Drive.
   • Action 4: delete the temporary bucket object.
   • Action 5: log the entire transaction to your audit store.
4. Test the flow: Use a non‑sensitive test document, verify that the temporary file disappears, and confirm that logs contain the expected fields.
5. Review permissions quarterly: Ensure the service account still only has drive.file access and that bucket policies have not broadened.

Following this checklist lets a solo founder or a small team automate summarization without exposing raw customer data to the LLM or to unnecessary Google Drive permissions.

When to involve AISecAll

If you need a security review of your Make scenario, custom encryption scripts, or help setting up tamper‑evident logging, AISecAll offers a short‑term audit service that fits the pace of small‑business projects.