{
 "public_id": "AIR-2025-0152",
 "incident_id": "2025-10-brave-unseeable-screenshot-injection",
 "date_disclosed": "2025-10-21",
 "date_occurred": null,
 "title": "Unseeable prompt injections in screenshots: Comet OCR injection and Fellou navigation injection",
 "class": "research_demo",
 "attack_surface": "browser_agent",
 "initial_vector": "screenshot_ocr_injection",
 "agent_capability_abused": [
  "screenshot_capture",
  "ocr",
  "page_navigation",
  "tool_use"
 ],
 "target_product": "Perplexity Comet; Fellou; Opera Neon",
 "target_vendor": "Perplexity; Fellou; Opera",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "data_exfiltration",
  "unauthorized_actions"
 ],
 "blast_radius": "users of the affected agentic browsers who screenshot or navigate to attacker pages",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI02"
  ],
  "mitre_atlas": [],
  "labels": [
   "multimodal_injection",
   "invisible_text",
   "ocr_injection",
   "navigation_injection",
   "cross_vendor"
  ]
 },
 "defense_present": "vendor content filters; user-visible action logs",
 "defense_bypassed": true,
 "fix_shipped": "partial; Opera Neon details released 2025-10-31 after coordination",
 "primary_sources": [
  "https://brave.com/blog/unseeable-prompt-injections/"
 ],
 "verified_quote": "Agentic browser assistants can be prompt-injected by untrusted webpage content, rendering protections such as the same-origin policy irrelevant because the assistant executes with the user's authenticated privileges.",
 "confidence": "high",
 "notes": "Two distinct techniques: (1) Comet - faint light-blue text on yellow background invisible to humans but recovered by OCR when the user screenshots the page; (2) Fellou - merely asking the browser to navigate to an attacker URL sends page text to the LLM where it overrides user intent. A fourth affected browser was withheld. Brave's stated conclusion: agentic browsing is inherently dangerous today.",
 "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 agents acted on OCR'd screenshot injections",
 "vector_family": "other",
 "vector_detail": "screenshot_ocr_injection",
 "guardrail_kind": [
  "unknown"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "perplexity_comet",
 "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": [
  "unsafe_actuation"
 ],
 "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"
}