vLLM before 0.28.0 Denial of Service via audio extraction
Key facts
- Attack vector
- Local
- Privileges required
- None
- User interaction
- None
- Attack complexity
- Low
- Attack requirements
- None
- Impact type
- Denial of service
- Confidentiality · integrity · availability
- None · None · High
- Fix status
- Partial Red Hat: 3 without a fix
Attack profile from CVSS 4.0 · CNA. Fix status from CNA and Red Hat. Impact type from CWE-400.
Summary
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.
Why it matters now
SSVC decision points
CISA ADP · SSVC 2.0.3
- Exploitation
- None
- Automatable
- No
- Technical impact
- Partial
The decision (Track, Attend, Act) also depends on your mission and well-being impact; see CISA's SSVC guide.
Affected products
| Vendor | Product | Affected versions | Fixed in | Status | Source |
|---|---|---|---|---|---|
| vllm-project | vLLM | 0.10.2 to before 0.28.0 | 0.28.0 | Patch available | CNA |
- vLLM Vendorvllm-project Affected 0.10.2 to before 0.28.0 Fixed in 0.28.0 Patch available CNA
Remediation
- Fix · CNA vLLM (vllm-project) Fixed in 0.28.0
Fixed in the same release
Records fixed by the same vendor advisory or release, as the CNA lists them.
- Release ·
vllm-project vLLM 0.28.0
3 records Highest priority Priority LOW 3 LOW
Other records, highest priority first
- TI-2026-61567 vLLM before 0.28.0 Remote Code Execution via LlavaOnevision2 processorPriority LOW 39
- TI-2026-61569 vLLM before 0.28.0 Denial of Service via Audio HeaderPriority LOW 32
Vendor assessments
Red Hat
Red Hat rating: Moderate. Red Hat's page for this CVE →
| Product | Status | Plan or fix |
|---|---|---|
| Red Hat AI Inference Server | Affected | No fix yet: Fix deferred |
| Red Hat Enterprise Linux AI (RHEL AI) 3 | Affected | No fix yet: Fix deferred |
| Red Hat OpenShift AI (RHOAI) | Affected | No fix yet: Fix deferred |
Timeline
- 2026-10-02 Threatist record created from CVE-2026-90554
- 2026-10-03 Alias added: GHSA-m39m-2pqx-3vm2
References
More vllm-project vulnerabilities
- TI-2024-10229 Remote Code Execution in vllm-project/vllm CVSS 3.0CRITICAL 9.8
- TI-2024-08404 Remote Code Execution in vllm-project/vllm CVSS 3.1CRITICAL 9.8
- TI-2026-36017 vLLM: OpenAI auth bypass CVSS 3.1CRITICAL 9.1
- TI-2026-63056 vLLM before 0.28.0 Denial of Service via negative token ID CVSS 4.0HIGH 8.7
- TI-2026-04308 Hardcoded trust_remote_code=True in vllm-project/vllm Bypasses User Security Control CVSS 3.0HIGH 8.8
- TI-2026-39137 vLLM: Remote DoS in vLLM via Invalid Recovered Token Reinjection CVSS 3.1HIGH 7.5
Sources
- CVE-2026-90554 · CVE Program
- CVE-2026-90554 · NVD
- FIRST EPSS
- GHSA-m39m-2pqx-3vm2 · GitHub Advisory Database · CC BY 4.0
- CVE-2026-90554 · Red Hat security data · CC BY 4.0
This product uses the NVD API but is not endorsed or certified by the NVD. Sources and licences