AIKIDO-2026-816969
vllm is vulnerable to Race Condition
53
Medium Risk
This Affects:
vllmTL;DR
Serialized prompt-embedding reconstruction is guarded by a process-global PyTorch sparse-invariant context rather than a request-local one. When two prompt-embedding parts from a single chat request are loaded concurrently on the default executor, one context can restore the global flag off while the other is still inside its guard. A malformed sparse payload can then bypass the invariant check and reach the dense-conversion sink, risking an engine crash. The fix serializes access so the shared invariant state cannot be disabled by concurrent requests.
Who does this affect?
You are affected if you are using a version that falls within the vulnerable range and you run with prompt embeddings enabled and expose the chat completions endpoint to untrusted clients.
Background info
vllm is vulnerable to Race Condition in versions 0.21.0 - 0.25.1.
How to fix this
Upgrade the vllm library to the patch version.
Are You Affected?
Connect your repositories to instantly see whether vulnerable or malicious packages exist in your codebase.
Free. No credit card required.
I consent to receiving marketing communications based on Aikido’s Privacy Policy.
SOC 2Compliant
ISO 27001Compliant