Disclosed vulnerability
Notion AI: document edits are saved before user approval, so a markdown-image injection exfiltrates data whether or not the user accepts
AIR-2026-0005 AIR record · 2026-01-notion-ai-preapproval-exfiltration
From the supporting source
Notion AI was susceptible to data exfiltration via indirect prompt injection due to a vulnerability in which AI document edits are saved before user approval.promptarmor.com
The record
Disclosed
2026-01-07
Vendor
Notion
Product
Notion AI
Risk domain
Privacy › Data exfiltration channels
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — Notion AI saved the edit and fetched the exfiltration image pre-approval
Attack surface
Enterprise assistant
Autonomy level
Assistant
Initial vector
Indirect prompt injection document
Impact type
Data exfiltration
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
Notion AI users processing untrusted documents (e.g. candidate resumes)
Capability abused
Defense & remediation
Recorded defense
user approval prompt for AI document edits
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
remediated 2026-01-07/08
Taxonomy & references
Curator notes
PromptArmor. Reported 2025-12-24, disclosed and fixed 2026-01-07. Demo scenario: a poisoned resume PDF leaks a hiring tracker (salary expectations, candidate feedback, diversity goals). The core defect is an ordering bug in the approval flow - 'Whether or not the user accepts the edit, the attacker successfully exfiltrates the data.' Distinct from 2025-09-notion-lethal-trifecta (Code Integrity, Notion 3.0 agents). | Merged from duplicate id `2026-01-notion-ai-unpatched-exfiltration`. Also reported as: 'Notion AI writes and saves exfiltration URLs into documents before user approval'.
Cite as
Enkrypt AI. "Notion AI: document edits are saved before user approval, so a markdown-image injection exfiltrates data whether or not the user accepts". AIR, AIR-2026-0005. https://agentincident.vercel.app/i/AIR-2026-0005/ (accessed 2026-10-19).