vulns.co
/
GKData.io MCP and API Buy me a coffee

Google · AI integration boundaries

Gemini-to-Colab rendering boundary exposed Workspace data

Content that Gemini treated as inactive could become active when exported to Colab. The researcher reports Workspace-data disclosure after a user exported content, in a scenario requiring access to connected data and exposure to untrusted content. Google awarded USD 20,000 for this finding in a multi-finding article; suggested delivery through training data or embeddings was not separately demonstrated.

Read the original writeupUSD 20,000 reported award · Reviewed 2026-10-03

Root cause

Content considered inert by Gemini could acquire active rendering behavior in Colab. The integration did not preserve the same sanitization contract across that boundary.

Demonstrated impact

The researcher reports confirming Workspace-data disclosure after a user exported content to Colab. The scenario depended on Gemini encountering untrusted content and having access to connected data. Suggested delivery through poisoned training data or embeddings was not separately demonstrated.

What to check for

  • Maintain consistent content-handling contracts across integration boundaries.
  • Enforce output destinations independently from generated or retrieved text.
  • Treat export as a new interpretation boundary; validate the destination representation rather than assuming upstream sanitization remains effective.

Award and evidence

USD 20,000Bug Bounty · Researcher Reported With Vendor Quote
How the award was verified, and what the record does not establish

This amount belongs to the Colab finding; a separate Gemini-only finding was marked duplicate.

  • Award correspondence is researcher-published; settlement is not independently verified.
  • The separate Gemini-only finding was a duplicate, not another award.
  • No exact publication day, fix date or vendor patch design is established.

Timeline

Published
2025-11explicit · Author homepage supplies November 2025; exact publication day is not established.
Reported
2025-04-30explicit
Awarded
2025-05-20explicit

Related visual models

Related cases

All 5 cases in AI integration boundaries

Sources

  1. Hacking Gemini: A Multi-Layered ApproachValentino Massaro · reviewed 2026-10-03
  2. Valentino’s issue trackerValentino Massaro · reviewed 2026-10-03

Record reviewed 2026-10-03. Snapshot af2809c308fc. Open the complete JSON contract.

Library synced 2026-11-18 · cases last reviewed 2026-10-05

af2809c308fc · JSON exports & schemas · CC BY 4.0 content / MIT software