Denial of Service Risk in vLLM Inference Engine by Vendor vLLM
CVE-2026-54233
6.5MEDIUM
What is CVE-2026-54233?
The vLLM inference and serving engine for large language models contains a vulnerability that allows for a denial of service. Prior to version 0.23.1rc0, the /v1/audio/transcriptions endpoint did not effectively limit the size of the decoded PCM output, which could expand significantly from a compressed OPUS file. As a result, a file as small as 25MB could decode into roughly 14.9GB of float32 PCM data, potentially overwhelming system resources and leading to service interruptions for users. This issue has been addressed in subsequent updates to ensure safer handling of audio uploads.
Affected Version(s)
vllm < 0.23.1rc0
