{
 "id": "AIR-2026-0618",
 "slug": "vllm-speculative-decoding-crashes-server-via-penalty-paramet",
 "title": "vLLM speculative decoding crashes server via penalty parameters",
 "summary": "A vulnerability in vLLM allows a denial of service attack by sending requests with specific sampling penalty parameters. The issue causes the inference engine to crash deterministically.",
 "type": "attack_on_ai",
 "categories": [
  "infra_vuln"
 ],
 "severity": 2,
 "targets": [
  "framework_or_tooling"
 ],
 "products": [
  "vLLM"
 ],
 "incident_date": "2026-05-06",
 "incident_date_precision": "day",
 "first_reported": "2026-05-06",
 "sources": [
  {
   "url": "https://github.com/advisories/GHSA-83vm-p52w-f9pw",
   "publisher": "GitHub Advisory Database",
   "title": "CVE-2026-44223: vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters",
   "published_at": "2026-05-06"
  },
  {
   "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-44223",
   "publisher": "NVD",
   "title": "CVE-2026-44223: vLLM is an inference and serving engine for large language models (LLMs).",
   "published_at": "2026-05-12"
  }
 ],
 "source_count": 2,
 "confidence": 0.95,
 "status": "published",
 "origin": "seed_ghsa",
 "tags": [
  "CVE-2026-44223",
  "dos",
  "vllm"
 ],
 "created_at": "2026-10-18T07:00:14.101Z",
 "updated_at": "2026-10-18T10:04:21.885Z"
}