{
 "public_id": "AIR-2023-0017",
 "incident_id": "2023-07-openai-custom-instructions-backdoor",
 "date_disclosed": "2023-07-24",
 "date_occurred": null,
 "title": "ChatGPT Custom Instructions used as a persistent data exfiltration backdoor",
 "class": "research_demo",
 "attack_surface": "consumer_chatbot",
 "initial_vector": "prompt_injection_custom_instructions",
 "agent_capability_abused": [
  "custom_instruction_persistence",
  "markdown_image_rendering"
 ],
 "target_product": "ChatGPT Custom Instructions",
 "target_vendor": "OpenAI",
 "autonomy_level": "assistant",
 "impact_realized": false,
 "impact_type": [
  "data_exfiltration",
  "persistence"
 ],
 "blast_radius": "ChatGPT users whose custom instructions were modified",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI03"
  ],
  "mitre_atlas": [],
  "labels": [
   "persistence",
   "image_markdown_exfiltration",
   "post_exploitation"
  ]
 },
 "defense_present": "none",
 "defense_bypassed": true,
 "fix_shipped": "not fixed at publication; the underlying image-markdown exfiltration was reported to OpenAI in April 2023",
 "primary_sources": [
  "https://embracethered.com/blog/posts/2023/chatgpt-custom-instruction-post-exploitation-data-exfiltration/"
 ],
 "verified_quote": "More and more plugins are taking a dependency on this insecure behavior for rendering images for instance, which will make a fix quite difficult for OpenAI the longer they wait.",
 "confidence": "high",
 "notes": "Johann Rehberger. First persistence primitive in ChatGPT, predating the memory feature.",
 "risk_category": "security_fraud",
 "risk_secondary": [
  "privacy"
 ],
 "risk_subcategory": "Prompt Injection",
 "risk_sub_secondary": [
  "Supply Chain Compromise",
  "Persistence & Covert C2"
 ],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": false,
 "agency_reason": "the assistant rendered exfiltration images from persisted custom instructions",
 "vector_family": "indirect_prompt_injection",
 "vector_detail": "prompt_injection_custom_instructions",
 "guardrail_kind": [
  "none"
 ],
 "guardrail_outcome": "not_applicable",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "unknown",
 "model_provider": "openai",
 "model_name": "",
 "detection_mechanism": "external_researcher",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "privilege_escalation",
 "eai_domain": "access_control_and_permissions",
 "eai_secondary": [
  "data_exfiltration_channels"
 ],
 "eai_reason": "impact rule: privilege_escalation",
 "asi_curated": [
  "ASI01"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI01 | added ASI03 (eai_category=privilege_escalation, no code execution)",
 "submitted_by": "EAI Internal"
}