CVE-2026-100651
Received Received - Intake

Buffer Overflow in vLLM Disaggregated Serving Endpoint

Publication date: 2026-09-26

Last updated on: 2026-09-26

Assigner: VulnCheck

Description
vLLM before 0.29.0 fails to enforce decoder prompt-length validation on the disaggregated serving endpoint /inference/v1/generate. When the request contains a 'features' (multimodal) payload, vllm/entrypoints/serve/disagg/serving.py builds a multimodal EngineInput directly from the caller-supplied token_ids, and GenerateRequest.token_ids (vllm/entrypoints/serve/disagg/protocol.py) is not checked against model_config.max_model_len. For multimodal processors that report skip_prompt_length_check=True (for example Nemotron Parse, Whisper, and FireRedLID), InputProcessor._validate_prompt_len() returns immediately for both encoder and decoder prompts, so an overlong prompt becomes an EngineCoreRequest and reaches the worker input-batch copy into a fixed max_model_len-wide NumPy row. A client able to reach the endpoint on an affected model configuration can therefore submit an overlong token_ids list to trigger a worker failure and denial of service. Fixed in 0.29.0.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published
2026-09-26
Last Modified
2026-09-26
Generated
2026-09-26
AI Q&A
2026-09-26
EPSS Evaluated
N/A
NVD
CVE-2026-100651
EUVD
EUVD-2026-87745

Affected Vendors & Products

Showing 1 associated CPE
Vendor Product Version / Range
vllm_project vllm to 0.29.0 (exc)

Helpful Resources

Exploitability

CWE
CWE Icon
KEV
KEV Icon
CWE ID Description
CWE-400 The product does not properly control the allocation and maintenance of a limited resource.

Attack-Flow Graph

AI Quick Actions

Instant insights powered by AI
Executive Summary

This vulnerability affects vLLM versions before 0.29.0. It involves a failure to enforce decoder prompt-length validation on the disaggregated serving endpoint /inference/v1/generate when handling multimodal payloads. Certain multimodal processors (Nemotron Parse, Whisper, FireRedLID) skip prompt-length checks, allowing attackers to submit overlong token_ids lists. This bypasses validation and causes worker failures during input batching, leading to denial of service.

Detection Guidance

Check if your vLLM version is below 0.29.0 by running: vllm --version. Inspect disaggregated serving endpoints for /inference/v1/generate. Monitor for failed worker processes or input batching errors during multimodal requests.

Impact Analysis

An attacker could exploit this to cause a denial of service by submitting an excessively long decoder prompt. This would crash the vLLM worker process during input batching, disrupting service availability for legitimate users. The impact is limited to availability and requires API access to the vulnerable endpoint.

Mitigation Strategies

Upgrade vLLM to version 0.29.0 or later. If upgrading is not possible, restrict access to the /inference/v1/generate endpoint to trusted users only. Disable multimodal features if not required.

Chat Assistant

Ask questions about this CVE
Hi! I’m here to help you understand CVE-2026-100651. Ask me anything about the vulnerability, its impact, or mitigation strategies.
0/70

EPSS Chart