cve.bz

CVE-2026-100653 vllm vllm: Use of Less Trusted Source

8.3 highExploit referenceEPSS 0.3%Fix available

High (8.3) in vllm vllm. EPSS 0.3% estimated probability of exploitation activity in the next 30 days. Fixed in vllm 0.28.0.

On this page
vLLM is an inference and serving engine for large language models. In versions from 0.22.1 through 0.28.0, the operator-supplied model revision pin (--revision / --code-revision) is not propagated to several Hugging Face artifact loads for the FunAudioChat and Tarsier2 architectures: the WhisperFeatureExtractor and speech_tokenizer PreTrainedTokenizerFast loads in vllm/model_executor/models/funaudiochat.py and the Qwen2VLConfig.from_pretrained call used by Tarsier2ProcessingInfo in vllm/model_executor/models/qwen2_vl.py. As a result, deployments pinned to a reviewed revision still resolve these behavior-affecting processor, tokenizer, and config artifacts from the repository's default revision, so a later change to the upstream default branch can alter audio preprocessing, speech tokenizer behavior, or Tarsier2 configuration without any change to the operator's configured pin. This is a supply-chain integrity and reproducibility failure for pinned deployments; it is residual to the earlier fix tracked as GHSA-3ww4-5jv9-j5gm / CVE-2026-47155 and does not constitute remote code execution or a trust_remote_code=False bypass. The issue is fixed in version 0.28.0.

Severity

CVSS 4.0

8.3 high
CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:L/VI:H/VA:N/SC:N/SI:N/SA:N

CVSS 3.1

6.5 medium
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:H/A:N
EPSS
0.3% (unchanged from 0.3% a week ago; 0.3% 30 days ago) estimated probability of exploitation activity in the next 30 days; 28th percentile. Scored 2026-11-17. A forecast for this CVE across the internet, not for your own systems.
CISA's assessment (SSVC)
Exploitation: PoC · Automatable: no · Technical impact: partial, assessed 2026-09-30. From CISA's Vulnrichment in the CVE List: whether it is exploited, whether exploiting it can be automated, and how much control it gives.
CWE
CWE-348 Use of Less Trusted Source
Published
Modified
NVD analysis
Analyzed
Record from
the NVD, synced

Fix and patched versions

Ask an AI about this CVE

Show the promptExplain CVE-2026-100653 for an engineer deciding whether to act. Use https://cve.bz/api/public/cves/CVE-2026-100653?via=ask-ai-copy and https://cve.bz/cve/CVE-2026-100653?via=ask-ai-copy as the source of truth: what is affected and which versions fix it, whether it is in CISA KEV, public exploits, and its EPSS (an internet-wide 30-day estimate of exploitation activity, not a forecast for my systems). If they give no fixed versions, use the CVE record https://cveawg.mitre.org/api/cve/CVE-2026-100653 or the vendor advisory, not third-party mirrors. Tell me how to check whether I am affected and how to verify a fix. Cite the sources and say where they are uncertain.

Opens a third-party assistant; its answer is its own, not cve.bz's. The prompt holds only this CVE's id and public links.

Affected

Versions

From the CNA's record in the CVE List.

ProductAffected
vllm-project vllmfrom 0.22.1 before 0.28.0

Packages (OSV)

AdvisoryEcosystemPackageAffectedFixed in
PYSEC-2026-4324PyPIvllm>=0.22.1 <0.28.00.28.0

Red Hat

Red Hat rates this Moderate.

Red Hat's statement

Red Hat is not impacted. Our portfolio makes use of vllm versions ranging from 0.8.5-0.24.0.

As per the validated model matrix in ai-inference-server-3.2/3.3/3.4 (rhaiis/rhaii vllm-*) it has no FunAudioChat or Tarsier2 instead it's (Llama/Granite/Qwen/Phi/etc only). Images can still run customer-supplied Hugging Face models but that is Bring Your Own and not product default.

rhel-ai-3.4 (bootc-*-rhel9) focuses on Hugging Face local Granite-class models and not FunAudioChat/Tarsier2. There is no FunAudioChat/Tarsier2 serving in rhoai-* odh-kserve-{agent,controller,router} as well. Also, rhoai-* odh-kserve-storage-initializer (includes vulnerable vllm version 0.24.0) but, does not run FunAudioChat/Tarsier2 inference paths.

rhoai-* odh-vllm-gaudi is Gaudi vLLM runtime and still not a FunAudioChat/Tarsier2 validated model. Vllm versions here are also outside CVE range (0.14.1 / 0.19.1). rhoai-3.4/3.5 odh-llm-d-kv-cache is KV-cache component and not those model architectures. Lastly rhoai-3.5 odh-th-torch-cuda-py312 which is torch/training image with vLLM dep 0.21.0 which is also outside the published CVE range.

Red Hat's mitigation

Pre-download the vetted model repository to local storage and configure offline execution.

1. Download the specific, verified commit revision of the model to a local directory:

huggingface-cli download <MODEL_NAME> --revision <REVISION_HASH> --local-dir /path/to/local/model

2. Configure the environment to disable remote Hub requests and launch vLLM using the local path:

export HF_HUB_OFFLINE=1

vllm serve /path/to/local/model

For containerized environments (such as Red Hat OpenShift AI or Red Hat Enterprise Linux AI), set the environment variable HF_HUB_OFFLINE=1 in the container deployment specification and mount the pre-downloaded model volume into the container filesystem.

Caveats:

Deployments require sufficient local or persistent volume storage to host model weights and configurations.

Warning: Modifying the deployment configuration or environment variables requires restarting the serving container or service, resulting in temporary inference downtime.

ProductPackageStateAdvisory
Red Hat AI Inference Serverrhaiis/vllm-cpu-rhel9Fix deferred
Red Hat AI Inference Serverrhaiis/vllm-cuda-rhel9Fix deferred
Red Hat AI Inference Serverrhaiis/vllm-neuron-rhel9Fix deferred
Red Hat AI Inference Serverrhaiis/vllm-rocm-rhel9Fix deferred
Red Hat AI Inference Serverrhaiis/vllm-spyre-rhel9Fix deferred
Red Hat AI Inference Serverrhaiis/vllm-tpu-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-cpu-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-cuda-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-gaudi-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-neuron-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-rocm-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-spyre-rhel9Fix deferred
Red Hat AI Inference Serverrhaii/vllm-tpu-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-aws-cuda-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-azure-cuda-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-azure-rocm-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-cuda-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-gaudi-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-gcp-cuda-rhel9Fix deferred
Red Hat Enterprise Linux AI (RHEL AI) 3rhelai3/bootc-rocm-rhel9Fix deferred
7 more in the JSON and on Red Hat's page.

Contains Red Hat Security Data © Red Hat, Inc., licensed under CC BY 4.0. Source: https://access.redhat.com/security/data. Modified: reformatted. Red Hat's page for CVE-2026-100653.

Attack patterns (CAPEC)

  • CAPEC-73 User-Controlled Filename Likelihood: High · Severity: High
  • CAPEC-76 Manipulating Web Input to File System Calls Likelihood: High · Severity: Very High
  • CAPEC-85 AJAX Footprinting Likelihood: High · Severity: Low
  • CAPEC-141 Cache Poisoning Likelihood: High · Severity: High
  • CAPEC-142 DNS Cache Poisoning Likelihood: High · Severity: High

The attack patterns MITRE's CAPEC relates to this record's weaknesses, as MITRE names them. Credits.

References

Check your stackJSON