{
 "public_id": "AIR-2025-0146",
 "incident_id": "2025-10-unit42-bedrock-memory-poisoning",
 "date_disclosed": "2025-10-09",
 "date_occurred": null,
 "title": "Persistent memory poisoning of Amazon Bedrock Agents via indirect prompt injection",
 "class": "research_demo",
 "attack_surface": "memory_store",
 "initial_vector": "indirect_prompt_injection_tool_output",
 "agent_capability_abused": [
  "long_term_memory",
  "orchestration_prompt_assembly",
  "retrieval"
 ],
 "target_product": "Amazon Bedrock Agents (memory feature)",
 "target_vendor": "Amazon Web Services",
 "autonomy_level": "assistant",
 "impact_realized": false,
 "impact_type": [
  "data_exfiltration",
  "persistence"
 ],
 "blast_radius": "Bedrock Agent deployments with memory enabled",
 "cve": [],
 "taxonomy": {
  "owasp_asi": [
   "ASI01",
   "ASI06",
   "ASI02"
  ],
  "mitre_atlas": [],
  "labels": [
   "memory_poisoning",
   "cross_session_persistence"
  ]
 },
 "defense_present": "Amazon Bedrock Guardrails prompt-attack policy (not enabled in the tested configuration)",
 "defense_bypassed": false,
 "fix_shipped": "No code fix; AWS points customers to Bedrock Guardrails",
 "primary_sources": [
  "https://unit42.paloaltonetworks.com/indirect-prompt-injection-poisons-ai-longterm-memory/"
 ],
 "verified_quote": "Once planted, these instructions persist across sessions, and they are incorporated into the agent's orchestration prompts. This ultimately allows the agent to silently exfiltrate a user's conversation history in future interactions.",
 "confidence": "high",
 "notes": "Authors Royce Lu and Jay Chen. Amazon responded that the concerns are mitigable by enabling Bedrock platform features. Proof of concept, not observed in the wild.",
 "risk_category": "security_fraud",
 "risk_secondary": [
  "privacy"
 ],
 "risk_subcategory": "Prompt Injection",
 "risk_sub_secondary": [
  "Persistence & Covert C2"
 ],
 "ai_entity": "generative",
 "agency": "agent_acted",
 "agency_debatable": false,
 "agency_reason": "the Bedrock agent wrote poisoned memory and exfiltrated on later turns",
 "vector_family": "indirect_prompt_injection",
 "vector_detail": "indirect_prompt_injection_tool_output",
 "guardrail_kind": [
  "input_classifier"
 ],
 "guardrail_outcome": "not_enabled",
 "xrefs": {
  "aiid": [],
  "atlas": [],
  "cve": [],
  "ghsa": []
 },
 "primary_asi": "ASI01",
 "agent_framework": "unknown",
 "model_provider": "unknown",
 "model_name": "",
 "detection_mechanism": "external_researcher",
 "tool_access": "unknown",
 "financial_loss_band": "unknown",
 "eai_category": "data_and_memory_poisoning",
 "eai_domain": "reliability_and_observability",
 "eai_secondary": [
  "privilege_escalation",
  "data_exfiltration_channels",
  "confused_deputy"
 ],
 "eai_reason": "impact rule: data_and_memory_poisoning",
 "asi_curated": [
  "ASI01"
 ],
 "asi_scope": "in_scope",
 "asi_provenance": "curated_plus_rule",
 "asi_reason": "curated: ASI01 | added ASI06 (eai_category=data_and_memory_poisoning), ASI02 (impact=tool_misuse)",
 "submitted_by": "EAI Internal"
}