{
 "public_id": "AIR-2026-0208",
 "incident_id": "2026-08-anthropic-claudebleed-reopened",
 "date_disclosed": "2026-08-28",
 "date_occurred": null,
 "title": "ClaudeBleed reopened: any extension can dispatch a synthetic click to make Claude for Chrome run tasks against Gmail, Drive and Salesforce",
 "class": "responsible_disclosure",
 "attack_surface": "browser_agent",
 "initial_vector": "synthetic_dom_event_from_a_co_resident_extension",
 "agent_capability_abused": [
  "task_execution",
  "authenticated_session_reuse",
  "email_read",
  "calendar_write",
  "crm_write"
 ],
 "target_product": "Claude for Chrome",
 "target_vendor": "Anthropic",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "data_exfiltration",
  "unauthorized_actions"
 ],
 "blast_radius": "Claude for Chrome users with any other extension that has content-script access to claude.ai",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI04",
   "ASI03"
  ],
  "mitre_atlas": [],
  "labels": [
   "untrusted_event",
   "missing_isTrusted_check",
   "incomplete_fix",
   "confused_deputy",
   "extension_permission_model"
  ]
 },
 "defense_present": "user click required to launch a task (not verified as a trusted event)",
 "defense_bypassed": true,
 "fix_shipped": "none as of disclosure - Anthropic closed the reports and marked the tracking issue Resolved; the handler in v1.0.80 is byte-identical to v1.0.72",
 "primary_sources": [
  "https://www.manifold.security/blog/claude-for-chrome-extension-bypass"
 ],
 "verified_quote": "Any script with DOM access on claude.ai can construct the button, set the task ID, and dispatch a synthetic click.",
 "confidence": "high",
 "notes": "Manifold Security. Follow-up to their earlier 'ClaudeBleed' disclosure, which Anthropic had announced fixed. The click handler never checks event.isTrusted, so a co-resident extension fabricates the click and Claude executes a chosen task ID with the user's live sessions - demonstrated reading Gmail and Google Docs comments, creating Calendar meetings, and modifying Salesforce leads. Researchers note the fix is one line. DISPUTE: vendor considers it resolved; researchers say it is not.",
 "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": "a synthetic click makes the agent run tasks against Gmail, Drive and Salesforce",
 "vector_family": "other",
 "vector_detail": "synthetic_dom_event_from_a_co_resident_extension",
 "guardrail_kind": [
  "unknown"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI04",
 "agent_framework": "unknown",
 "model_provider": "anthropic",
 "model_name": "",
 "detection_mechanism": "unknown",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "data_exfiltration_channels",
 "eai_domain": "privacy",
 "eai_secondary": [
  "unsafe_actuation",
  "confused_deputy"
 ],
 "eai_reason": "subcategory: Data Exfiltration via Context",
 "asi_curated": [
  "ASI04"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI04 | added ASI03 (signal=identity)",
 "submitted_by": "EAI Internal"
}