CVE-2026-114052: vLLM memory exhaustion in MOSS-Audio model serving

CVE-2026-114052 · Severity: high · CVSS 7.5 · Published 2026-10-21

vLLM is a library that serves large language and multimodal models. An attacker can exploit the MOSS-Audio model serving feature by sending specially crafted requests to the /tokenize endpoint, which bypasses API key authentication, to exhaust the server's memory and crash the API process. This creates a denial-of-service vulnerability that affects any deployment running MOSS-Audio models.

Technologies: vllm (PyPI), vLLM Project vllm. Vendors: PyPI, vLLM Project.

Quick answers

Is CVE-2026-114052 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-114052 being exploited in the wild.
How severe is CVE-2026-114052?
High severity, with a CVSS score of 7.5, published 21 October 2026.
Which products and versions are affected?
vllm (PyPI) ECOSYSTEM:introduced=0.24.0,fixed=0.31.0; vLLM Project vllm after 0.24.0 and before 0.31.0.
Is there a fix for CVE-2026-114052?
Yes. vllm (PyPI): upgrade to 0.31.0 or later. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-114052 attacked?
It is attacked over the network, in a software library.
Am I affected by CVE-2026-114052?
Check whether you run vllm (PyPI) or vLLM Project vllm at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

How to fix

  • vllm (PyPI): upgrade to 0.31.0 or later
    pip install "vllm>=0.31.0"

Technical details

The vulnerability is an unbounded resource allocation issue (CWE-770) in MOSS-Audio model serving where mm_processor_kwargs values from client requests are cached without limits in a private dictionary. Each distinct processor configuration instantiates a WhisperFeatureExtractor that allocates a mel-filter array; attackers can trigger persistent linear memory growth by submitting distinct configurations or amplify single-request allocations with oversized mel_config values. The /tokenize endpoint is intentionally unauthenticated and outside API-key-protected route prefixes, enabling unauthenticated remote exploitation.

Affected products

  • vLLM Project vLLM after 0.24.0 and before 0.31.0

Identifiers

  • CVE-2026-114052
  • GHSA-823j-m4cj-hmjf
  • PYSEC-2026-4361

Timeline

  • 2026-10-06: disclosed

References

This page as Markdown. Free to reuse under CC BY 4.0 with a link back.