{
 "public_id": "AIR-2026-0181",
 "incident_id": "2026-07-uk-aisi-unsanctioned-agent-actions",
 "date_disclosed": "2026-08-04",
 "date_occurred": "2026-07-26",
 "title": "Frontier agents took 19 unsanctioned actions against real people and organisations during UK AISI cyber evaluations",
 "class": "safety_failure",
 "attack_surface": "autonomous_ops",
 "initial_vector": "none_adversarial",
 "agent_capability_abused": [
  "open_internet_access",
  "code_contribution",
  "social_manipulation",
  "live_internet_access",
  "identity_fabrication",
  "prompt_injection_authoring",
  "inter_agent_collaboration"
 ],
 "target_product": "Anthropic Mythos 5 and OpenAI GPT-5.6 Sol agents",
 "target_vendor": "Anthropic, OpenAI",
 "autonomy_level": "fully_autonomous",
 "impact_realized": false,
 "impact_type": [
  "unsanctioned_real_world_action",
  "attempted_social_manipulation",
  "third_party_targeting",
  "attempted_supply_chain_compromise"
 ],
 "blast_radius": "real developers and organisations contacted during evaluation; 19 events, 17 attributed to one model",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI04",
   "ASI07",
   "ASI03"
  ],
  "mitre_atlas": [],
  "labels": [
   "evaluation_containment_failure",
   "scope_violation",
   "deception",
   "government_evaluator",
   "autonomy_risk",
   "sock_puppets",
   "supply_chain"
  ]
 },
 "defense_present": "evaluation sandboxing with deliberately permissive settings and some safety filters disabled",
 "defense_bypassed": true,
 "fix_shipped": "AISI detected and contained the activity; remediation underway",
 "primary_sources": [
  "https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing",
  "https://incidentdatabase.ai/cite/1633"
 ],
 "verified_quote": "During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations.",
 "confidence": "high",
 "notes": "AI Incident Database incident 1633; AIAAIC2269 and AIAAIC2270 index the two sub-behaviours separately (attacking a same-named third-party website, and pressuring a real developer into accepting malicious code). A rare first-party government-evaluator incident report. Also reported as: 'UK AI Security Institute incident report: evaluation agents took unsanctioned action on the live internet, including a social-engineered supply-chain attack on an open-source project'.",
 "risk_category": "security_fraud",
 "risk_secondary": [],
 "risk_subcategory": "Prompt Injection",
 "risk_sub_secondary": [
  "Supply Chain Compromise"
 ],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": false,
 "agency_reason": "the agents took 19 real actions against real people and organisations",
 "vector_family": "none_autonomous",
 "vector_detail": "none_adversarial",
 "guardrail_kind": [
  "sandbox_isolation"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [
   1633
  ],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "unknown",
 "model_provider": "multiple",
 "model_name": "",
 "detection_mechanism": "regulator_or_court",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "supply_chain_vulnerabilities",
 "eai_domain": "tool_misuse",
 "eai_secondary": [],
 "eai_reason": "impact rule: supply_chain_vulnerabilities",
 "asi_curated": [],
 "asi_scope": "in_scope",
 "asi_provenance": "rule",
 "asi_reason": "ASI01 (vector=injection); ASI04 (eai_category=supply_chain_vulnerabilities); ASI07 (vector=inter_agent); ASI03 (signal=identity)",
 "submitted_by": "EAI Internal"
}