All vulnerabilities

Servers and applications

GHSA-3mqx-f33v-vgp9

Disaggregated generate skips decoder prompt-length validation for some multimodal processors

Severity
Medium 6.5
Published
12 Sept 2026
Last changed
12 Sept 2026
Added to UpHeron
3 Nov 2026
Fixed in
0.29.0
Weakness
Uncontrolled resource consumption (CWE-400)Improper check for unusual conditions (CWE-754)
Attack
Over the network · Needs a user account · No user action needed
Attacker can
Take it down
CVSS vector
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Sources
Project advisories (GitHub)

Description

The Python /inference/v1/generate disaggregated serving endpoint has a features path that builds a multimodal EngineInput from caller-supplied token_ids:

That EngineInput is not an EngineCoreRequest; AsyncLLM.add_request() routes it through InputProcessor.process_inputs() before scheduling. For ordinary processors, this is a safe blocking layer: _validate_prompt_len() rejects decoder prompts whose length is greater than max_model_len before constructing an EngineCoreRequest.

The reachable issue is narrower. InputProcessor sets a process-wide skip_prompt_length_check flag from the active multimodal processor:

When that flag is true, _validate_prompt_len() immediately returns for both encoder and decoder prompts. Current head includes processors that return True for this property, including Nemotron Parse, Whisper, and FireRedLID.

From vllm-project/vllm's security advisory. Shortened; the full text is there.

Affected software

SoftwareAffected versions
vLLMServers and applications · github:vllm-project/vllm
  • Every version before 0.29.0

About vLLM

open-source software for large language model inference

From Wikidata, CC0 1.0.
Docker Hub pulls
44,195,225Docker Hub, 15 Nov 2026
GitHub stars
96,943GitHub, 15 Nov 2026
Forks
24,836GitHub, 15 Nov 2026
Latest release
v0.33.0rc3GitHub, 15 Nov 2026
Released
2 Nov 2026GitHub
License
Apache Software License 2.0Wikidata
First published on GitHub
9 Feb 2023GitHub

References