CVE-2026-105922: vLLM denial of service in token binning penalty handler

CVE-2026-105922 · Severity: medium · CVSS 4.3 · Published 2026-10-06

vLLM is a high-throughput inference engine for large language models. A flaw in its penalty handling mechanism can crash the inference engine when processing requests with certain prompt embeddings and token penalties, causing service unavailability. The issue occurs with concurrent requests and has a public exploit available.

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

Quick answers

Is CVE-2026-105922 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-105922 being exploited in the wild.
How severe is CVE-2026-105922?
Medium severity, with a CVSS score of 4.3, published 6 October 2026.
Which products and versions are affected?
vllm (PyPI) ECOSYSTEM:introduced=0,fixed=0.32.0; vLLM Project vllm up to 0.31.0.
Is there a fix for CVE-2026-105922?
Yes. vllm (PyPI): upgrade to 0.32.0 or later. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-105922 attacked?
It is attacked over the network, in a software library.
Am I affected by CVE-2026-105922?
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 0.32.0 or later
    pip install "vllm>=0.32.0"

Technical details

The vulnerability exists in the get_token_bin_counts_and_mask function within vllm/model_executor/layers/utils.py, which is part of the penalty handler component. An attacker can cause a device-side assertion failure (scatter-gather kernel index out of bounds) by sending concurrent requests that combine prompt embeddings with token repetition, presence, or frequency penalties. Exploitation requires network access to the vLLM inference endpoint.

Affected products

  • vllm-project vLLM up to 0.31.0

Identifiers

  • CVE-2026-105922
  • PYSEC-2026-4319

Timeline

  • 2026-10-06: disclosed
  • exploited: public exploit released

References

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