{
 "public_id": "AIR-2025-0053",
 "incident_id": "2025-07-supabase-mcp-sql-exfiltration",
 "date_disclosed": "2025-07-06",
 "date_occurred": null,
 "title": "Supabase MCP can leak your entire SQL database: a support-ticket injection makes the developer's agent run a service_role query that bypasses RLS",
 "class": "research_demo",
 "attack_surface": "mcp_server",
 "initial_vector": "indirect_prompt_injection_support_ticket",
 "agent_capability_abused": [
  "sql_execution",
  "service_role_credentials",
  "ticket_reply_write"
 ],
 "target_product": "Supabase MCP server",
 "target_vendor": "Supabase",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "data_exfiltration"
 ],
 "blast_radius": "Supabase projects whose developers run the MCP server with service_role credentials",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI02"
  ],
  "mitre_atlas": [],
  "labels": [
   "lethal_trifecta",
   "rls_bypass",
   "service_role_overprivilege",
   "support_ticket_injection"
  ]
 },
 "defense_present": "Postgres Row-Level Security (bypassed by service_role)",
 "defense_bypassed": true,
 "fix_shipped": "no patch; mitigations are read-only mode and prompt injection filters",
 "primary_sources": [
  "https://www.generalanalysis.com/blog/supabase-mcp-blog"
 ],
 "verified_quote": "The core problem of LLMs interacting with tools is that they can not distinguish instructions from data.",
 "confidence": "high",
 "notes": "General Analysis. DATE CAVEAT: Simon Willison linked and covered this on 2025-07-06, which is used as date_disclosed; the page itself now shows 'Published April 10, 2026 (Reviewed May 1, 2026)', apparently a re-dating of the article. The stolen integration_tokens are unreachable to the human support agent's own credentials - the escalation comes entirely from the developer's MCP session running as service_role. Canonical worked example of Willison's lethal trifecta.",
 "risk_category": "privacy",
 "risk_secondary": [],
 "risk_subcategory": "Data Exfiltration via Context",
 "risk_sub_secondary": [],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": false,
 "agency_reason": "the agent ran the service_role query and wrote the data into a public reply",
 "vector_family": "indirect_prompt_injection",
 "vector_detail": "indirect_prompt_injection_support_ticket",
 "guardrail_kind": [
  "unknown"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "mcp_server",
 "model_provider": "unknown",
 "model_name": "",
 "detection_mechanism": "external_researcher",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "data_exfiltration_channels",
 "eai_domain": "privacy",
 "eai_secondary": [
  "confused_deputy"
 ],
 "eai_reason": "subcategory: Data Exfiltration via Context",
 "asi_curated": [
  "ASI01"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI01 | added ASI02 (impact=tool_misuse)",
 "submitted_by": "EAI Internal"
}