Skip to content
Fencer Vulns
medium 6.5

CVE-2026-44223 json

vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters

published 2026-05-06 6 mo ago modified 2026-09-10 1 mo ago also EUVD-2026-29800 , GHSA-83vm-p52w-f9pw , PYSEC-2026-145

cvss
6.5
version 3.1, CNA
cisa kev
Not listed
not in CISA's catalog
epss
0.4%
EPSS 0.4% chance of exploitation in 30 days, percentile 30%
public exploits
0
none known

Summary

vLLM is an inference and serving engine for large language models (LLMs). From 0.18.0 to before 0.20.0, the extract_hidden_states speculative decoding proposer in vLLM returns a tensor with an incorrect shape after the first decode step, causing a RuntimeError that crashes the EngineCore process. The crash is triggered when any request in the batch uses sampling penalty parameters (repetition_penalty, frequency_penalty, or presence_penalty). A single request with a penalty parameter (e.g., "repetition_penalty": 1.1) is sufficient to crash the server. This vulnerability is fixed in 0.20.0.

Patch / fix

  • pypi vllm 0.20.0
  • vllm-project vllm 0.20.0
  • vllm vllm 0.20.0

Vendor advisories

What the vendors themselves say about their products: affected, fixed or not affected, and what to do.

red hat product security vex CVE-2026-44223 updated cvss 7.5

vllm: vLLM: Denial of Service via malformed tensor shape in speculative decoding

  • Red Hat AI Inference Server rhaii/vllm-tpu-rhel9, rhaiis/vllm-cpu-rhel9, rhaiis/vllm-tpu-rhel9, rhaii/vllm-gaudi-rhel9, rhaiis/vllm-cuda-rhel9, rhaiis/vllm-rocm-rhel9, +3 more
    affected: workaround, no fix available, no fix planned “Will not fix”
  • Red Hat OpenShift AI (RHOAI) rhoai/odh-vllm-gaudi-rhel9
    affected: workaround, no fix planned “Will not fix”
  • Red Hat AI Inference Server 3.4 rhaii/vllm-cpu-rhel9 1787151769, rhaii/vllm-cpu-rhel9 1789681128, rhaii/vllm-cpu-rhel9 1790075793, rhaii/vllm-cuda-rhel9 1787151771, rhaii/vllm-cuda-rhel9 1789681126, rhaii/vllm-cuda-rhel9 1790090131, +6 more
    fixed: vendor fix RHSA-2026:57380
  • Red Hat Enterprise Linux AI 3.4 rhelai3/bootc-cuda-rhel9 1787243961, rhelai3/bootc-rocm-rhel9 1787244006, rhelai3/bootc-aws-cuda-rhel9 1787253912, rhelai3/bootc-gcp-cuda-rhel9 1787253774, rhelai3/disk-image-cuda-rhel9 1787310717, rhelai3/bootc-azure-cuda-rhel9 1787253989, +1 more
    fixed: vendor fix RHSA-2026:59151
  • Red Hat OpenShift AI (RHOAI) rhoai/odh-vllm-cuda-rhel9, rhoai/odh-vllm-rocm-rhel9, rhoai/odh-kserve-agent-rhel9, rhoai/odh-kserve-router-rhel9, rhoai/odh-kserve-controller-rhel9, rhoai/odh-kserve-storage-initializer-rhel9
    not affected: vulnerable code not present

Red Hat CSAF VEX © Red Hat, Inc., from https://security.access.redhat.com/data/csaf/v2/vex/; changed: reduced to product statuses and remediations. CC BY 4.0 · csaf

CISA SSVC

Exploitation
Proof of concept is public
Automatable
no
Technical impact
partial
Assessed
5 mo ago

Exploit prediction

0.37% chance of exploitation in the next 30 days, higher than 30% of all scored CVEs (FIRST EPSS, 2026-11-08).

Exploits and checks

  • github.com https://github.com/vllm-project/vllm/pull/38610 referenced PoC, tagged "exploit" by a source, unverified
  • github.com https://github.com/vllm-project/vllm/security/advisories/GHSA-83vm-p52w-f9pw referenced PoC, tagged "exploit" by a source, unverified

CVSS scores

assigned by version score vector
CNA 3.1 6.5 CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
GitHub 3.1 6.5 CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
OSV 3.1 6.5 CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Affected packages

ecosystem package affected fixed in
pypi vllm
>=0.18.0, <0.20.0
0.20.0

Affected products

  • vllm-project vllm

Classification

Weakness
  • CWE-131 Incorrect Calculation of Buffer Size
  • CWE-704 Incorrect Type Conversion or Cast

References

Sources

Fencer Vulns keeps one entry per vulnerability: the CVE id wins, then GHSA, and records that alias each other are merged, except that two CVEs always keep separate entries. The same data is available as JSON.

CPEs and fixed versions use Anchore's correction of NVD's data: Added CPE configurations because not yet analyzed by NVD.