{
  "id": "CVE-2026-90554",
  "url": "https://poc.griffinguard.io/cve/CVE-2026-90554",
  "published_at": "2026-09-12T12:08:57.508Z",
  "title": "vLLM before 0.28.0 Denial of Service via audio extraction",
  "vendor": "vllm-project",
  "product": "vLLM",
  "cvss": {
    "score": 7.5,
    "source": "nvd",
    "vector": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H"
  },
  "kev": {
    "in_kev": false,
    "added_at": null,
    "days_after_publication": null
  },
  "first_poc": null,
  "poc_repo_count": 0,
  "exploit_count": 0,
  "suspicious_repo_count": 0,
  "edge_device": null,
  "description": "vLLM versions \u003e=0.10.2 and \u003c0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.",
  "cwe": "CWE-400",
  "cpes": [
    "cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*"
  ],
  "cvss_by_source": {
    "cna": 6.9,
    "nvd": 7.5
  },
  "repos": [],
  "remediation": {
    "affected": [
      {
        "vendor": "vllm-project",
        "product": "vLLM",
        "ranges": [
          "\u003e= 0.10.2, \u003c 0.28.0"
        ],
        "fixed_in": [
          "0.28.0"
        ]
      }
    ],
    "upgrade": [
      "Upgrade vLLM to 0.28.0 or later."
    ],
    "references": [
      {
        "kind": "Vendor advisory",
        "url": "https://github.com/vllm-project/vllm/security/advisories/GHSA-936p-m5pv-vvjf"
      }
    ]
  },
  "exploit_refs": [],
  "hunt_profile": [],
  "rules": []
}
