{
 "public_id": "AIR-2025-0066",
 "incident_id": "2025-07-grayswan-agent-red-teaming-arena",
 "date_disclosed": "2025-07-28",
 "date_occurred": null,
 "title": "Agent Red Teaming competition: 1.8M prompt-injection attacks break all 22 frontier agents on the Gray Swan Arena",
 "class": "research_demo",
 "attack_surface": "other",
 "initial_vector": "direct_and_indirect_prompt_injection",
 "agent_capability_abused": [
  "tool_use",
  "retrieval",
  "code_execution",
  "computer_use"
 ],
 "target_product": "Gray Swan Arena (22 frontier AI agents across 44 deployment scenarios)",
 "target_vendor": "Gray Swan AI; UK AI Security Institute; participating frontier labs",
 "autonomy_level": "semi_autonomous",
 "impact_realized": false,
 "impact_type": [
  "policy_violation",
  "data_exfiltration",
  "unauthorized_action"
 ],
 "blast_radius": "22 frontier agents; 44 realistic deployment scenarios",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI05"
  ],
  "mitre_atlas": [],
  "labels": [
   "competition",
   "crowdsourced_red_team",
   "universal_transfer",
   "art_benchmark"
  ]
 },
 "defense_present": "vendor-native agent safety training and system prompts",
 "defense_bypassed": true,
 "fix_shipped": null,
 "primary_sources": [
  "https://arxiv.org/abs/2507.20526"
 ],
 "verified_quote": "Participants submitted 1.8 million prompt-injection attacks, with over 60,000 successfully eliciting policy violations",
 "confidence": "high",
 "notes": "Zou, Lin, Jones, Nowak, Dziemian, Davies, Kirk, Gal, Hendrycks, Kolter, Fredrikson et al. Title: 'Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition'; targets '22 frontier AI agents across 44 realistic deployment scenarios'. Produced the ART benchmark. Agent identities were anonymised during the competition, so per-model attribution comes from the paper rather than the live arena.",
 "risk_category": "security_fraud",
 "risk_secondary": [
  "privacy"
 ],
 "risk_subcategory": "Prompt Injection",
 "risk_sub_secondary": [
  "Remote Code Execution & Sandbox Escape"
 ],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": true,
 "agency_reason": "the 22 agents took the tool, retrieval and code-execution actions in scenario",
 "vector_family": "indirect_prompt_injection",
 "vector_detail": "direct_and_indirect_prompt_injection",
 "guardrail_kind": [
  "prompt_instruction",
  "abuse_detection"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "unknown",
 "model_provider": "unknown",
 "model_name": "",
 "detection_mechanism": "academic_study",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "unsafe_actuation",
 "eai_domain": "agent_behaviour",
 "eai_secondary": [
  "data_exfiltration_channels",
  "confused_deputy"
 ],
 "eai_reason": "impact rule: unsafe_actuation",
 "asi_curated": [
  "ASI01"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI01 | added ASI05 (vector=code_exec)",
 "submitted_by": "EAI Internal"
}