CVE-2026-57173
Published:September 16, 2026
Updated:September 22, 2026
vLLM is an inference and serving engine for large language models. Prior to 0.24.0, the input_audio handling path for /v1/chat/completions calls AudioMediaIO.load_bytes or AudioMediaIO.load_file without passing VLLM_MAX_AUDIO_DECODE_DURATION_S to the shared audio decoder. An unauthenticated client can therefore submit a small compressed audio input that expands into a very large float32 PCM allocation, bypassing the duration guard already used by /v1/audio/transcriptions and causing an out-of-memory worker crash. Inline data URLs reach this path without being bounded by VLLM_AUDIO_FETCH_TIMEOUT. The issue affects deployments serving an audio-capable model, and authentication changes only the deployment-specific reachability. This issue is fixed in version 0.24.0.
Affected Packages
vllm (CONDA):
Affected version(s) >=0.8.3 <0.24.0Fix Suggestion:
Update to version 0.24.0https://github.com/vllm-project/vllm.git (GITHUB):
Affected version(s) >=v0.1.0 <v0.24.0Fix Suggestion:
Update to version v0.24.0vllm (PYTHON):
Affected version(s) >=0.0.1 <0.24.0Fix Suggestion:
Update to version 0.24.0Related Resources (6)
Do you need more information?
Contact UsCVSS v4
Base Score:
7.1
Attack Vector
NETWORK
Attack Complexity
LOW
Attack Requirements
NONE
Privileges Required
LOW
User Interaction
NONE
Vulnerable System Confidentiality
NONE
Vulnerable System Integrity
NONE
Vulnerable System Availability
HIGH
Subsequent System Confidentiality
NONE
Subsequent System Integrity
NONE
Subsequent System Availability
NONE
CVSS v3
Base Score:
6.5
Attack Vector
NETWORK
Attack Complexity
LOW
Privileges Required
LOW
User Interaction
NONE
Scope
UNCHANGED
Confidentiality
NONE
Integrity
NONE
Availability
HIGH
Weakness Type (CWE)
Allocation of Resources Without Limits or Throttling
EPSS
Base Score:
0.66