Disclosed vulnerability
ChatGPT for Google Sheets: one injection yields Apps Script execution, cross-workbook exfiltration and a phishing overlay, ignoring the user's approval setting
AIR-2026-0122 AIR record · 2026-05-openai-chatgpt-google-sheets-apps-script
From the supporting source
This attack does not require human-in-the-loop approvals, even when in settings the user has explicitly required human approvalpromptarmor.com
The record
Disclosed
2026-05-27
Vendor
OpenAI
Product
ChatGPT for Google Sheets extension
Risk domain
Privacy › Data exfiltration channels
Also touches
Credential theft, Confused deputy
System type
Generative model
Who acted
Agent acted — the extension executed Apps Script and exfiltrated across workbooks
Attack surface
Enterprise assistant
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection spreadsheet
Impact type
Data exfiltration, Credential theft, Unauthorized modification
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
users of the ChatGPT for Google Sheets extension with connectors or imported sheets
Capability abused
Defense & remediation
Recorded defense
user setting explicitly requiring human approval
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
OpenAI removed the model's ability to generate Apps Script code (2026-05-31) and is re-evaluating sandboxing
Taxonomy & references
Curator notes
PromptArmor. Disclosed 2026-05-08, public 2026-05-27, OpenAI responded 2026-05-31. The extension's Apps Script capability runs with the OAuth permissions the user granted the extension, so a single injection reaches every workbook in the account.
Cite as
Enkrypt AI. "ChatGPT for Google Sheets: one injection yields Apps Script execution, cross-workbook exfiltration and a phishing overlay, ignoring the user's approval setting". AIR, AIR-2026-0122. https://agentincident.vercel.app/i/AIR-2026-0122/ (accessed 2026-10-19).