{
 "id": "AIR-2026-1402",
 "slug": "stealthy-poisoning-attacks-via-reused-kv-cache-of-reasoning",
 "title": "Stealthy Poisoning Attacks via Reused KV Cache of Reasoning",
 "summary": "Researchers demonstrate a black-box poisoning attack called R2ECAP that plants poisoned reasoning traces by exploiting reused key-value caches in large language models. The attack manipulates model behavior solely through interactions with the chat interface.",
 "type": "attack_on_ai",
 "categories": [
  "data_poisoning"
 ],
 "severity": 1,
 "targets": [
  "open_weight_model"
 ],
 "products": [],
 "incident_date": "2026-10-19",
 "incident_date_precision": "day",
 "first_reported": "2026-10-19",
 "sources": [
  {
   "url": "https://arxiv.org/abs/2610.22873",
   "publisher": "arXiv",
   "title": "Stealthy Poisoning Attacks via Reused KV Cache of Reasoning",
   "published_at": "2026-10-19"
  }
 ],
 "source_count": 1,
 "confidence": 0.95,
 "status": "published",
 "origin": "news",
 "tags": [
  "kv-cache",
  "data-poisoning",
  "reasoning-traces",
  "black-box-attack"
 ],
 "created_at": "2026-10-19T05:07:38.588Z",
 "updated_at": "2026-10-19T05:07:38.588Z"
}