Out-of-Memory Vulnerability in vLLM for Audio Processing
CVE-2026-57173

6.5MEDIUM

Key Information:

Status
Vendor
CVE Published:
16 September 2026

What is CVE-2026-57173?

A vulnerability exists in vLLM, an inference engine for large language models, affecting audio processing capabilities. Prior to version 0.24.0, an unauthenticated client could manipulate the audio input handling for /v1/chat/completions. This occurs because the system fails to enforce limits on the audio decoding duration. Consequently, a small compressed audio input could be exploited to expand into an excessively large float32 PCM allocation, leading to an out-of-memory crash of the worker. This issue can be triggered by inline data URLs that bypass the VLLM_AUDIO_FETCH_TIMEOUT limit, putting deployments that serve audio-capable models at risk. The vulnerability is addressed in version 0.24.0.

Affected Version(s)

vllm < 0.24.0

References

CVSS V3.1

Score:
6.5
Severity:
MEDIUM
Confidentiality:
None
Integrity:
None
Availability:
None
Attack Vector:
Network
Attack Complexity:
Low
Privileges Required:
Low
User Interaction:
None
Scope:
Unchanged

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.