CVE-2026-90554: vLLM denial of service through unbounded audio decoding from video input

Summary

vLLM versions 0.10.2 up to but not including 0.28.0 apply no audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. The _extract_audio_from_videos function in nano_nemotron_vl.py calls load_audio_pyav without the max_duration_s or max_decode_bytes parameters, so VLLM_MAX_AUDIO_DECODE_DURATION_S and VLLM_MAX_AUDIO_DECODE_BYTES are not enforced. When such a model is served with use_audio_in_video=True, an attacker supplying a small, highly compressed video can make the server allocate gigabytes of memory during audio decoding, causing a denial of service.

Mitigation

Fixed in vLLM 0.28.0.

Topics

• Inference infrastructure

Related items

• LowCloakFEM: Agentic FEM Code Generation for Red-Teaming Computational Resource ExhaustionSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• MediumThe Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model ResourcesSimilar attack · Google Threat Intelligence
• InfoThe Blind Spots of Detection: Adversarial Perturbation Attacks against Watchdogs in Distributed LLM InferenceSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• HighCVE-2026-118761: Kubeflow Pipelines decompression bomb in pipeline archive uploadSimilar attack · NVD/CVE Database
• LowOpenCode User Hit with Hostile Backlash After Reporting Free AI LoopholeSimilar attack · Socket Blog