{
 "id": "AIR-2026-0237",
 "slug": "vllm-flash-late-interaction-scoring-permits-cross-request-in",
 "title": "vLLM flash late-interaction scoring permits cross-request integrity break",
 "summary": "Caller-controlled request identifiers in vLLM flash late-interaction scoring permit query embeddings to be cached improperly. This creates cross-request integrity issues in affected deployments.",
 "type": "attack_on_ai",
 "categories": [
  "infra_vuln"
 ],
 "severity": 2,
 "targets": [
  "framework_or_tooling"
 ],
 "products": [
  "vLLM"
 ],
 "incident_date": "2026-10-05",
 "incident_date_precision": "day",
 "first_reported": "2026-10-05",
 "sources": [
  {
   "url": "https://github.com/advisories/GHSA-2phq-3phc-84px",
   "publisher": "GitHub Advisory Database",
   "title": "CVE-2026-105755: vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and…",
   "published_at": "2026-10-05"
  },
  {
   "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-105755",
   "publisher": "NVD",
   "title": "CVE-2026-105755: vLLM is an inference and serving engine for large language models.",
   "published_at": "2026-10-05"
  }
 ],
 "source_count": 2,
 "confidence": 0.95,
 "status": "published",
 "origin": "seed_ghsa",
 "tags": [
  "CVE-2026-105755",
  "cache-poisoning",
  "vllm"
 ],
 "created_at": "2026-10-18T06:33:31.967Z",
 "updated_at": "2026-10-18T06:33:31.967Z"
}