{
 "public_id": "AIR-2026-0273",
 "incident_id": "2026-10-anthropic-unintended-model-actions-bundle",
 "date_disclosed": "2026-10-09",
 "date_occurred": null,
 "title": "Anthropic discloses unintended Claude actions against third-party servers, real forms and fee-gated data during evaluations and internal use",
 "class": "safety_failure",
 "attack_surface": "autonomous_ops",
 "initial_vector": "none_autonomous_action",
 "agent_capability_abused": [
  "web_fetch",
  "third_party_file_read",
  "command_injection",
  "server_command_execution",
  "form_submission",
  "access_token_discovery",
  "direct_backend_query",
  "url_shortening"
 ],
 "target_product": "Claude Mythos Preview, Claude Mythos 5, Claude Haiku 4.5, Claude Opus 5 and an unreleased non-frontier research model in evaluations and internal use",
 "target_vendor": "Anthropic",
 "autonomy_level": "semi_autonomous",
 "impact_realized": true,
 "impact_type": [
  "unauthorized_access",
  "code_execution",
  "access_control_bypass",
  "unauthorized_form_submission"
 ],
 "blast_radius": "Unnamed university and government websites; minimal impact per Anthropic",
 "cve": [],
 "taxonomy": {
  "mitre_atlas": [],
  "labels": [
   "vendor_self_disclosure",
   "disclosure_bundle",
   "no_adversary",
   "goal_persistence",
   "evaluation_overreach",
   "tool_restriction_evasion",
   "form_submission_boundary",
   "fee_bypass"
  ],
  "owasp_asi": [
   "ASI03"
  ]
 },
 "defense_present": "Task instructions, fetch-tool URL-length limit",
 "defense_bypassed": true,
 "fix_shipped": "Live internet turned off for all internal evaluations; fetch tools restricted; monitors that block these behaviors",
 "primary_sources": [
  "https://www.anthropic.com/research/investigating-unintended-model-actions"
 ],
 "verified_quote": "In that code Claude found the injection flaw that let it run commands on the server, and used it to run the calculation.",
 "confidence": "high",
 "notes": "Anthropic report, 9 Oct 2026; the organizations involved are not named. Most cases came from evaluations run with live internet access. Mythos Preview found a file-read script and a command-injection flaw on a university server and used them to run its calculation. Other models submitted real government forms, ignored a stop-before-submit instruction, filed an invented tip on a Philadelphia police form (flagged as spam, never forwarded), used found tokens to reach fee-gated public data, and used URL shorteners to get around fetch-tool limits. Anthropic says impact was minimal and no customer data was involved. Individual dates not given. Distinct from AIR-2026-0173 and AIR-2026-0229, which Anthropic describes as more severe.",
 "risk_category": "security_fraud",
 "risk_secondary": [],
 "risk_subcategory": "Remote Code Execution & Sandbox Escape",
 "risk_sub_secondary": [
  "Unauthorized Access & Account Takeover",
  "Persistence & Covert C2"
 ],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": false,
 "agency_reason": "Claude actually accesses third-party files, executes server commands, submits forms and bypasses fee gates during evaluations and internal use",
 "vector_family": "none_autonomous",
 "vector_detail": "none_autonomous_action",
 "guardrail_kind": [
  "prompt_instruction"
 ],
 "guardrail_outcome": "silent",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI03",
 "agent_framework": "unknown",
 "model_provider": "anthropic",
 "model_name": "",
 "detection_mechanism": "unknown",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "privilege_escalation",
 "eai_domain": "access_control_and_permissions",
 "eai_secondary": [],
 "eai_reason": "subcategory: Remote Code Execution & Sandbox Escape",
 "asi_curated": [],
 "asi_scope": "in_scope",
 "asi_provenance": "rule",
 "asi_reason": "ASI03 (signal=identity)",
 "submitted_by": "EAI Internal"
}