{
 "id": "AIR-2026-0240",
 "slug": "vllm-denial-of-service-via-unbounded-video-sampler-frames",
 "title": "vLLM Denial of Service via Unbounded Video Sampler Frames",
 "summary": "An unauthenticated caller can submit malicious video frame parameters to the vLLM tokenizer endpoint, consuming excessive frontend memory and crashing the API process. The vulnerability is addressed in updated framework releases.",
 "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://nvd.nist.gov/vuln/detail/CVE-2026-105758",
   "publisher": "NVD",
   "title": "CVE-2026-105758: vLLM is an inference and serving engine for large language models.",
   "published_at": "2026-10-05"
  },
  {
   "url": "https://github.com/advisories/GHSA-x6mc-67gf-chw4",
   "publisher": "GitHub Advisory Database",
   "title": "CVE-2026-105758: vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach",
   "published_at": "2026-10-05"
  }
 ],
 "source_count": 2,
 "confidence": 0.95,
 "status": "published",
 "origin": "seed_nvd",
 "tags": [
  "CVE-2026-105758",
  "denial-of-service",
  "vllm"
 ],
 "created_at": "2026-10-18T06:33:31.967Z",
 "updated_at": "2026-10-18T06:33:31.967Z"
}