{
 "public_id": "AIR-2024-0010",
 "incident_id": "2024-02-openai-gpts-code-interpreter-isolation",
 "date_disclosed": "2024-02-14",
 "date_occurred": null,
 "title": "ChatGPT: lack of isolation between Code Interpreter sessions of different GPTs",
 "class": "responsible_disclosure",
 "attack_surface": "skill_plugin",
 "initial_vector": "malicious_custom_gpt",
 "agent_capability_abused": [
  "code_execution",
  "sandbox_file_access"
 ],
 "target_product": "ChatGPT Code Interpreter / custom GPTs",
 "target_vendor": "OpenAI",
 "autonomy_level": "assistant",
 "impact_realized": false,
 "impact_type": [
  "unauthorized_data_access",
  "data_tampering"
 ],
 "blast_radius": "users who used multiple GPTs with Code Interpreter in the same session",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI04",
   "ASI05",
   "ASI03"
  ],
  "mitre_atlas": [],
  "labels": [
   "sandbox_isolation",
   "cross_tenant",
   "malicious_plugin"
  ]
 },
 "defense_present": "assumed per-GPT sandbox isolation, which did not exist",
 "defense_bypassed": true,
 "fix_shipped": "fixed May 2024; reported November 2023 with no follow-up before the 90-day public disclosure",
 "primary_sources": [
  "https://embracethered.com/blog/posts/2024/lack-of-isolation-gpts-code-interpreter/"
 ],
 "verified_quote": "It allows the creator of a malicious GPT to steal and also overwrite files from your conversations with other GPTs, including ChatGPT.",
 "confidence": "high",
 "notes": "Johann Rehberger. An isolation failure rather than a prompt-injection bug.",
 "risk_category": "privacy",
 "risk_secondary": [],
 "risk_subcategory": "Control bypass in normal operation",
 "risk_sub_secondary": [],
 "risk_override_reason": "one GPT's Code Interpreter session could reach another's files; the harm is cross-tenant data access, not fraud",
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": true,
 "agency_reason": "the Code Interpreter executed attacker code that read and overwrote other sessions' files",
 "vector_family": "malicious_component",
 "vector_detail": "malicious_custom_gpt",
 "guardrail_kind": [
  "sandbox_isolation"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI04",
 "agent_framework": "custom_gpt",
 "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": [
  "supply_chain_vulnerabilities"
 ],
 "eai_reason": "subcategory: Control bypass in normal operation",
 "asi_curated": [],
 "asi_scope": "in_scope",
 "asi_provenance": "rule",
 "asi_reason": "ASI04 (vector_family=malicious_component); ASI05 (vector=code_exec); ASI03 (signal=identity)",
 "submitted_by": "EAI Internal"
}