Research demo
Indirect prompt injection via YouTube transcripts against ChatGPT plugins
AIR-2023-0009 AIR record · 2023-05-openai-youtube-transcript-injection
From the supporting source
If ChatGPT accesses the transcript, the owner of the video/transcript takes control of the chat session and gives the AI a new identity and objective.embracethered.com
The record
Disclosed
2023-05-14
Vendor
OpenAI (third-party plugin)
Product
ChatGPT YouTube plugin
Risk domain
Privacy › Data exfiltration channels
Also touches
Human manipulation, Confused deputy
System type
Generative model
Who acted
Agent acted — the plugin retrieved the poisoned transcript and acted on it
Attack surface
Skill / plugin
Autonomy level
Assistant
Initial vector
Indirect prompt injection video transcript
Impact type
Session hijack, Scam, 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
ChatGPT users summarising untrusted videos
Capability abused
Defense & remediation
Recorded defense
none
Guardrail kinds
None
Guardrail outcome
Not applicable
Taxonomy & references
Curator notes
Johann Rehberger. Early proof that any user-generated content feed is an injection channel for tool-using assistants.
Supporting sources
Cite as
Enkrypt AI. "Indirect prompt injection via YouTube transcripts against ChatGPT plugins". AIR, AIR-2023-0009. https://agentincident.vercel.app/i/AIR-2023-0009/ (accessed 2026-10-19).