CVE-2026-100653: vLLM: revision pin bypass loads unpinned HF artifacts
MEDIUMCISA: TRACK*vLLM versions 0.22.1 through 0.28.0 fail to apply the operator's --revision / --code-revision pin to several Hugging Face loads for the FunAudioChat and Tarsier2 architectures, so processors, speech tokenizers and configs silently resolve from the repository's default branch. This is a supply-chain integrity gap, not remote code execution: a change upstream can alter audio preprocessing or model configuration without any change on your side, which undermines the reviewed-and-pinned assurance your change control assumes. Urgency is low to moderate, with a CVSS of 6.5 (high attack complexity), an EPSS of 0.00284 (81st percentile), no public exploit, no Nuclei template, no KEV listing and a CISA SSVC decision of TRACK*. The blast radius is wider than the severity suggests, because vLLM has 95 downstream dependents and 134 other CVEs on record, and the flaw only matters if you serve FunAudioChat or Tarsier2. Inventory any such deployments, upgrade to the fixed release (the advisory lists 0.28.0 as the fix, though the affected range is stated as ending at 0.28.0, so confirm the exact patched build against GHSA-hhv2-872h-628q), and mirror pinned artifacts into an internal registry with offline mode enabled.
What is the risk?
Medium overall. Exploitation needs a third party who can alter the upstream Hugging Face repository's default revision, or an attacker who has compromised it, which explains the high attack complexity. There is no code-execution path and no trust_remote_code bypass, so impact is limited to integrity (high) and low confidentiality exposure. Likelihood is low (EPSS 0.28%, no public exploit, not in KEV), but exposure is real for any organisation that relies on revision pinning as a control for compliance or reproducibility. Organisations serving only other architectures are not affected. Treat as TRACK* per SSVC: monitor and patch in the normal cycle.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| vLLM | pip | — | No patch |
Do you use vLLM? You're affected.
How severe is it?
What is the attack surface?
What should I do?
1 step1) Identify vLLM instances serving FunAudioChat or Tarsier2 models and check the installed version (pip show vllm). 2) Upgrade to the fixed release per GHSA-hhv2-872h-628q and the upstream commit d26a28ab; verify the exact patched version in the advisory, since the description lists 0.28.0 as both the end of the affected range and the fix. 3) Mirror reviewed model repositories, including processor, tokenizer and config files, into an internal registry or artifact store and serve them from local paths. 4) Set HF_HUB_OFFLINE=1 or TRANSFORMERS_OFFLINE=1 in production so nothing is resolved from the Hub at runtime. 5) Record file hashes of all loaded artifacts and compare them with the approved baseline at deploy time. 6) Monitor egress to huggingface.co from inference hosts and alert on unexpected fetches after startup.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-100653?
vLLM versions 0.22.1 through 0.28.0 fail to apply the operator's --revision / --code-revision pin to several Hugging Face loads for the FunAudioChat and Tarsier2 architectures, so processors, speech tokenizers and configs silently resolve from the repository's default branch. This is a supply-chain integrity gap, not remote code execution: a change upstream can alter audio preprocessing or model configuration without any change on your side, which undermines the reviewed-and-pinned assurance your change control assumes. Urgency is low to moderate, with a CVSS of 6.5 (high attack complexity), an EPSS of 0.00284 (81st percentile), no public exploit, no Nuclei template, no KEV listing and a CISA SSVC decision of TRACK*. The blast radius is wider than the severity suggests, because vLLM has 95 downstream dependents and 134 other CVEs on record, and the flaw only matters if you serve FunAudioChat or Tarsier2. Inventory any such deployments, upgrade to the fixed release (the advisory lists 0.28.0 as the fix, though the affected range is stated as ending at 0.28.0, so confirm the exact patched build against GHSA-hhv2-872h-628q), and mirror pinned artifacts into an internal registry with offline mode enabled.
Is CVE-2026-100653 actively exploited?
No confirmed active exploitation of CVE-2026-100653 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-100653?
1) Identify vLLM instances serving FunAudioChat or Tarsier2 models and check the installed version (pip show vllm). 2) Upgrade to the fixed release per GHSA-hhv2-872h-628q and the upstream commit d26a28ab; verify the exact patched version in the advisory, since the description lists 0.28.0 as both the end of the affected range and the fix. 3) Mirror reviewed model repositories, including processor, tokenizer and config files, into an internal registry or artifact store and serve them from local paths. 4) Set HF_HUB_OFFLINE=1 or TRANSFORMERS_OFFLINE=1 in production so nothing is resolved from the Hub at runtime. 5) Record file hashes of all loaded artifacts and compare them with the approved baseline at deploy time. 6) Monitor egress to huggingface.co from inference hosts and alert on unexpected fetches after startup.
What systems are affected by CVE-2026-100653?
This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference, multimodal inference, speech and audio pipelines, containerized ML deployments.
What is the CVSS score for CVE-2026-100653?
CVE-2026-100653 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.34%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0010.001AI SoftwareAML.T0010.003ModelAML.T0018Manipulate AI ModelCompliance Controls Affected
What are the technical details?
Original Advisory
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.
Exploitation Scenario
An organisation pins a reviewed revision of a FunAudioChat model on a vLLM server and treats the pin as its supply-chain control. An adversary who has gained commit rights to the upstream Hugging Face repository, for example through a compromised maintainer account, pushes a modified feature extractor or speech tokenizer configuration to the default branch. At the next restart or scale-out event the vLLM server fetches these unpinned artifacts, so audio inputs are preprocessed or tokenized differently and the model's behaviour changes subtly. The operator's pin and review records show nothing amiss, so the tampering goes undetected and degrades output integrity.
Weaknesses (CWE)
CWE-348 — Use of Less Trusted Source: The product has two different sources of the same data or information, but it uses the source that has less support for verification, is less trusted, or is less resistant to attack.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:H/A:NReferences
Timeline
Related Vulnerabilities
CVE-2026-6173210.0Decepticon: ChatML token injection yields sandbox RCE
Same package: vllmCVE-2026-259609.8vllm: SSRF allows internal network access
Same package: vllmCVE-2025-472779.8vLLM: RCE via exposed TCPStore in distributed inference
Same package: vllmCVE-2024-110419.8vllm: RCE via unsafe pickle deserialization in MessageQueue
Same package: vllmCVE-2024-90539.8vllm: RCE via unsafe pickle deserialization in RPC server
Same package: vllm