{
 "public_id": "AIR-2026-0120",
 "incident_id": "2026-05-microsoft-copilot-cowork-exfiltrates-files",
 "date_disclosed": "2026-05-26",
 "date_occurred": null,
 "title": "Microsoft Copilot Cowork: poisoned Skill makes the agent send itself a Teams message that leaks pre-authenticated SharePoint/OneDrive download links",
 "class": "responsible_disclosure",
 "attack_surface": "enterprise_assistant",
 "initial_vector": "malicious_skill",
 "agent_capability_abused": [
  "skill_execution",
  "graph_api_file_links",
  "teams_message_send",
  "email_send",
  "html_image_rendering"
 ],
 "target_product": "Microsoft Copilot Cowork (Microsoft 365)",
 "target_vendor": "Microsoft",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "data_exfiltration"
 ],
 "blast_radius": "M365 tenants where users can upload or share Copilot Cowork Skills",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI05",
   "ASI04",
   "ASI02"
  ],
  "labels": [
   "skill_poisoning",
   "approval_gap",
   "preauthenticated_link_leak",
   "html_image_exfil"
  ],
  "mitre_atlas": []
 },
 "defense_present": "human approval required for sensitive actions",
 "defense_bypassed": true,
 "fix_shipped": "unknown; mitigation is SharePoint download policy restriction",
 "primary_sources": [
  "https://www.promptarmor.com/resources/microsoft-copilot-cowork-exfiltrates-files"
 ],
 "verified_quote": "unlike other sensitive actions, sending emails and Teams messages to the active user does not require human approval",
 "confidence": "high",
 "notes": "PromptArmor. The approval carve-out for self-directed Teams/email messages is the pivot: the agent is steered into sending itself a message containing HTML image tags that fire network requests to the attacker when rendered, carrying Microsoft Graph pre-authenticated file download links for SharePoint/OneDrive documents. Linked from Simon Willison's prompt-injection and exfiltration-attacks tags. verified_quote is a verbatim clause from the page.",
 "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 sent itself a Teams message leaking pre-authenticated links",
 "vector_family": "malicious_component",
 "vector_detail": "malicious_skill",
 "guardrail_kind": [
  "human_approval"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "m365_copilot",
 "model_provider": "microsoft",
 "model_name": "",
 "detection_mechanism": "external_researcher",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "data_exfiltration_channels",
 "eai_domain": "privacy",
 "eai_secondary": [
  "supply_chain_vulnerabilities"
 ],
 "eai_reason": "subcategory: Data Exfiltration via Context",
 "asi_curated": [
  "ASI01",
  "ASI05"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI01,ASI05 | added ASI04 (vector_family=malicious_component), ASI02 (impact=tool_misuse)",
 "submitted_by": "EAI Internal"
}