{
 "public_id": "AIR-2025-0176",
 "incident_id": "2025-11-perplexity-comet-mcp-stdio-api",
 "date_disclosed": "2025-11-19",
 "date_occurred": null,
 "title": "Hidden chrome.perplexity.mcp.addStdioServer API in Comet let extensions run arbitrary local commands",
 "class": "responsible_disclosure",
 "attack_surface": "browser_agent",
 "initial_vector": "privileged_private_browser_api_reachable_from_web_origin",
 "agent_capability_abused": [
  "local_command_execution",
  "mcp_stdio_server_registration"
 ],
 "target_product": "Perplexity Comet",
 "target_vendor": "Perplexity",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "remote_code_execution",
  "malware_delivery",
  "data_exfiltration"
 ],
 "blast_radius": "Comet installs before the 2025-11-20 silent update",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI06",
   "ASI05"
  ],
  "mitre_atlas": [],
  "labels": [
   "sandbox_escape",
   "private_api",
   "extension_stomping",
   "ransomware_poc",
   "disputed"
  ]
 },
 "defense_present": "browser extension sandbox (bypassed by the vendor's own private API)",
 "defense_bypassed": true,
 "fix_shipped": "Perplexity silently disabled the MCP API on 2025-11-20 'out of an abundance of caution'",
 "primary_sources": [
  "https://www.helpnetsecurity.com/2025/11/20/perplexity-comet-browser-security-mcp-api/"
 ],
 "verified_quote": "If the industry doesn't establish boundaries now, we're setting a precedent where AI browsers can bypass decades of security principles",
 "confidence": "medium",
 "notes": "SquareX research; Help Net Security's coverage is used as the source in place of the researchers' own writeup at labs.sqrx.com (permitted fallback). Chain per SquareX: reach the perplexity.ai origin (XSS or MitM) or stomp the trusted Comet Analytics extension ID, then call the private API to register an arbitrary stdio MCP server = arbitrary local command execution; ransomware PoC shown. DISPUTED: Perplexity called the claims 'entirely false', saying the API needs developer mode, sideloading and user consent. Disclosed to Perplexity 2025-11-04.",
 "risk_category": "security_fraud",
 "risk_secondary": [
  "privacy"
 ],
 "risk_subcategory": "Remote Code Execution & Sandbox Escape",
 "risk_sub_secondary": [
  "Ransomware & Extortion"
 ],
 "ai_entity": "generative",
 "agency": "ai_as_target",
 "agency_debatable": true,
 "agency_reason": "a privileged private browser API reachable from any web origin",
 "vector_family": "other",
 "vector_detail": "privileged_private_browser_api_reachable_from_web_origin",
 "guardrail_kind": [
  "sandbox_isolation"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI06",
 "agent_framework": "unknown",
 "model_provider": "unknown",
 "model_name": "",
 "detection_mechanism": "unknown",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "privilege_escalation",
 "eai_domain": "access_control_and_permissions",
 "eai_secondary": [
  "data_exfiltration_channels"
 ],
 "eai_reason": "subcategory: Remote Code Execution & Sandbox Escape",
 "asi_curated": [
  "ASI06"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI06 | added ASI05 (impact=code_exec)",
 "submitted_by": "EAI Internal"
}