Research demo
Malicious ChatGPT GPTs quietly exfiltrate user data via image markdown
AIR-2023-0030 AIR record · 2023-12-openai-malicious-gpts-exfil
From the supporting source
A GPT can contain instructions that ask the user for information and simply send it elsewhere.embracethered.com
The record
Disclosed
2023-12-12
Vendor
OpenAI
Product
ChatGPT custom GPTs
Risk domain
Tool misuse › Supply chain vulnerabilities
Also touches
Credential theft, Data exfiltration channels
System type
Generative model
Who acted
Agent acted — the GPT rendered a markdown image that exfiltrated conversation data
Attack surface
Skill / plugin
Autonomy level
Assistant
Initial vector
Malicious custom gpt
Impact type
Data exfiltration, Credential harvesting
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI04 – Agentic Supply Chain Vulnerabilities
- MITRE ATLAS
- AML.T0040 – AI Supply Chain Compromise
- NIST AI RMF
- MAP 4.1
- EU AI Act
- Annex III §2
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool failure
- Mitigation
- Scan dependencies and container images for vulnerabilities; Use trusted sources.
- 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 third-party GPTs in the GPT Store
Capability abused
Defense & remediation
Recorded defense
OpenAI GPT publishing review
Guardrail kinds
Abuse detection
Guardrail outcome
Silent
Fix shipped
not fixed at publication - reported 2023-11-13, closed as Not Applicable on 2023-11-15
Taxonomy & references
Curator notes
Johann Rehberger. Bing Chat, Bard and Claude had already fixed equivalent image-render exfiltration at the time.
Supporting sources
Cite as
Enkrypt AI. "Malicious ChatGPT GPTs quietly exfiltrate user data via image markdown". AIR, AIR-2023-0030. https://agentincident.vercel.app/i/AIR-2023-0030/ (accessed 2026-10-19).