{
 "id": "AIR-2026-0810",
 "slug": "vllm-audio-downmixing-differential-enables-moderation-bypass",
 "title": "vLLM audio downmixing differential enables moderation bypass",
 "summary": "A processing differential in multi-channel audio downmixing enables hidden-input and moderation bypasses for audio models in vLLM. The vulnerability stems from a discrepancy between Librosa and standard weighting algorithms.",
 "type": "attack_on_ai",
 "categories": [
  "infra_vuln",
  "adversarial_evasion"
 ],
 "severity": 2,
 "targets": [
  "framework_or_tooling"
 ],
 "products": [
  "vLLM",
  "Librosa"
 ],
 "incident_date": "2026-07-17",
 "incident_date_precision": "day",
 "first_reported": "2026-04-02",
 "sources": [
  {
   "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-34760",
   "publisher": "NVD",
   "title": "CVE-2026-34760: vLLM is an inference and serving engine for large language models (LLMs).",
   "published_at": "2026-04-02"
  },
  {
   "url": "https://github.com/advisories/GHSA-6c4r-fmh3-7rh8",
   "publisher": "GitHub Advisory Database",
   "title": "CVE-2026-34760: vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio models",
   "published_at": "2026-07-17"
  }
 ],
 "source_count": 2,
 "confidence": 0.95,
 "status": "published",
 "origin": "seed_ghsa",
 "tags": [
  "CVE-2026-34760",
  "vllm",
  "librosa"
 ],
 "created_at": "2026-10-18T07:00:14.101Z",
 "updated_at": "2026-10-18T10:04:21.885Z"
}