Intel

AIKIDO-2026-889587

vllm is vulnerable to Denial of Service (DoS)

Denial of Service (DoS)GHSA-8737-qx52-hjff Published 2 days ago

43

Medium Risk

This Affects:

PYTHONvllm
0.24.0 - 0.25.1
Fixed in 0.26.0
Are you affected? Scan for Free

TL;DR

The derender endpoints accept caller-supplied response objects and detokenize every nested token-id list without enforcing model context length, output-token, choice-count, or response-size limits. A client can submit oversized nested token structures that the server decodes and buffers directly, consuming CPU and memory proportional to the attacker-chosen payload. This can degrade availability for other clients sharing the process. The fix validates derender payload sizes against the same bounds generation would enforce before decoding.

Who does this affect?

You are affected if you are using a version that falls within the vulnerable range and you expose the completions or chat-completions derender endpoints to API clients.

Background info

vllm is vulnerable to Denial of Service (DoS) in versions 0.24.0 - 0.25.1.

How to fix this

Upgrade the vllm library to the patch version.