{
 "public_id": "AIR-2026-0061",
 "incident_id": "2026-03-brave-agentic-oversharing-study",
 "date_disclosed": "2026-03-05",
 "date_occurred": null,
 "title": "Agentic oversharing: browser agents leak user chat history, email and personal data to e-commerce sites during ordinary shopping tasks",
 "class": "research_demo",
 "attack_surface": "browser_agent",
 "initial_vector": "no_adversary_agent_default_behavior",
 "agent_capability_abused": [
  "context_composition",
  "form_filling",
  "search_query_emission",
  "navigation_behavior"
 ],
 "target_product": "Browser-Use; AutoGen (with GPT-4o, o3, o4-mini backbones)",
 "target_vendor": "Browser Use; Microsoft (AutoGen); OpenAI (models)",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "privacy_violation",
  "data_exfiltration"
 ],
 "blast_radius": "users delegating shopping or form-filling to browser agents",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI05",
   "ASI02"
  ],
  "mitre_atlas": [],
  "labels": [
   "oversharing",
   "behavioral_leakage",
   "no_adversary",
   "prompt_level_guardrails_insufficient",
   "context_minimization"
  ]
 },
 "defense_present": "prompt-level privacy instructions",
 "defense_bypassed": true,
 "fix_shipped": "n/a - research recommends context minimization",
 "primary_sources": [
  "https://brave.com/blog/agentic-oversharing/"
 ],
 "verified_quote": "Simply instructing agents to be privacy-conscious at the prompt level is insufficient to address the depth and breadth of natural agentic oversharing.",
 "confidence": "high",
 "notes": "Brave research: 180 shopping tasks x 3 resource types (chat histories, emails, personal info) on Amazon and eBay = 1,080 runs. Two leakage dimensions - content (what the agent types) and behavior (what it clicks, searches and visits); behavioral oversharing consistently dominated. Removing task-irrelevant context improved task success by up to 17.9%, so privacy and utility aligned. Non-adversarial: the harm comes from the agent's own default behavior.",
 "risk_category": "privacy",
 "risk_secondary": [],
 "risk_subcategory": "PII Leakage",
 "risk_sub_secondary": [
  "Data Exfiltration via Context"
 ],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": false,
 "agency_reason": "the browser agents themselves sent chat history and email to shopping sites",
 "vector_family": "none_autonomous",
 "vector_detail": "no_adversary_agent_default_behavior",
 "guardrail_kind": [
  "prompt_instruction"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI05",
 "agent_framework": "unknown",
 "model_provider": "multiple",
 "model_name": "",
 "detection_mechanism": "external_researcher",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "sensitive_data_exposure",
 "eai_domain": "privacy",
 "eai_secondary": [
  "data_exfiltration_channels"
 ],
 "eai_reason": "subcategory: PII Leakage",
 "asi_curated": [
  "ASI05"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI05 | added ASI02 (impact=tool_misuse)",
 "submitted_by": "EAI Internal"
}