{
 "public_id": "AIR-2023-0004",
 "incident_id": "2023-02-microsoft-bing-chat-sydney-prompt-leak",
 "date_disclosed": "2023-02-08",
 "date_occurred": "2023-02-08",
 "title": "Bing Chat's system prompt (\"Sydney\") extracted by prompt injection during limited preview",
 "class": "responsible_disclosure",
 "attack_surface": "enterprise_assistant",
 "initial_vector": "direct_prompt_injection_untrusted_input",
 "agent_capability_abused": [
  "system_prompt_access",
  "retrieval"
 ],
 "target_product": "Bing Chat (New Bing)",
 "target_vendor": "Microsoft / OpenAI",
 "autonomy_level": "assistant",
 "impact_realized": false,
 "impact_type": [
  "information_disclosure"
 ],
 "blast_radius": "Bing Chat preview users; the confidential system prompt for all sessions",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI02"
  ],
  "mitre_atlas": [],
  "labels": [
   "prompt_injection",
   "system_prompt_extraction",
   "early_llm_product"
  ]
 },
 "defense_present": "confidentiality instruction in the system prompt",
 "defense_bypassed": true,
 "fix_shipped": "unknown",
 "primary_sources": [
  "https://arstechnica.com/information-technology/2023/02/ai-powered-bing-chat-spills-its-secrets-via-prompt-injection-attack/",
  "https://incidentdatabase.ai/cite/473"
 ],
 "verified_quote": "Early testers of Bing Chat successfully used prompt injection to reveal its built-in initial instructions, which contains a list of statements governing ChatGPT's interaction with users.",
 "confidence": "high",
 "notes": "Kevin Liu (Stanford) extracted the prompt days after launch. Foundational case: instruction-level confidentiality in an LLM product is not a security boundary. No third-party harm, hence impact_realized false. AI Incident Database incident 473.",
 "risk_category": "privacy",
 "risk_secondary": [],
 "risk_subcategory": "System Prompt Extraction",
 "risk_sub_secondary": [],
 "ai_entity": "generative",
 "agency": "elicitation_only",
 "agency_debatable": true,
 "agency_reason": "the model was made to emit its own system prompt; no victim, no action",
 "vector_family": "direct_prompt_injection",
 "vector_detail": "direct_prompt_injection_untrusted_input",
 "guardrail_kind": [
  "prompt_instruction"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [
   473
  ],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "unknown",
 "model_provider": "multiple",
 "model_name": "",
 "detection_mechanism": "unknown",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "sensitive_data_exposure",
 "eai_domain": "privacy",
 "eai_secondary": [],
 "eai_reason": "subcategory: System Prompt Extraction",
 "asi_curated": [],
 "asi_scope": "in_scope",
 "asi_provenance": "rule",
 "asi_reason": "ASI01 (vector_family=direct_prompt_injection); ASI02 (impact=tool_misuse)",
 "submitted_by": "EAI Internal"
}