CVE-2026-90554
vLLM versions >=0.10.2 and <0.28.0 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.
Scoring
- Severity
- MEDIUM
- CVSS base score
- 6.9
- CVSS vector
- CVSS:4.0/AV:L/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N
- EPSS probability
- 0.11%
- CWE
- CWE-400
- Published
- 2026-09-12
- Last modified
- 2026-09-14
Affected products
- vllm-project vLLM
- vllm-project vLLM
Weakness type
Related vulnerabilities
- CVE-2026-68763 — Apache Tomcat: DoS via allocation leak in HTTP/2 backlog tracking when a stream is reset
- CVE-2026-55247 — plone.app.event: Denial of service via iCalendar import
- CVE-2026-55248 — plone.app.portlets: Denial of service via RSS feed portlet
- CVE-2026-48050 — Arc: Unauthenticated access to Go debug pprof endpoints leaks runtime state and enables CPU-burn DoS
- CVE-2026-78551 — RansomLook Login Endpoint Allows Timing-Based Username Enumeration and Unthrottled Authentication Attempts
- CVE-2026-86452 — MISP Unauthenticated Mail Endpoints Allow Unbounded Storage Consumption and Request Flooding
- CVE-2026-86201 — PocketMine-MP before 5.41.1 LogDoS via LoginPacket clientData
- CVE-2026-85443 — MOOS core-moos through 10.4.0 MOOSDB Accept Loop Denial of Service