CVE-2026-105755

vLLM request identifiers could corrupt another scoring request

vllm-project · vLLM flash late-interaction scoring

First documented disclosure · Review checkpoint

Summary

A concurrent scoring request could replace the query embedding used for another request.

Severity and exploitation evidence

Medium · 4.2/10 · CVSS 3.1. Severity source.

Exploitation not established: Malicious exploitation remains unverified in the reviewed sources.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The worker’s query-cache key was derived from the caller-controlled request identifier.

Conditions required

The attacker has scoring API access, knows a victim’s request identifier, and wins a concurrency race on a flash late-interaction deployment.

Potential impact

Another request can receive incorrect scores or a transient cache-miss error. Data disclosure is not established.

Affected versions and configurations

Before 0.30.0.

Fix and mitigation status

0.30.0.

Practical next steps

Upgrade to 0.30.0 or later. Until patched, avoid shared flash late-interaction scoring for untrusted clients.

Operating lesson

Use server-generated request identities for shared inference caches.

Evidence and limits

The public correction PR opened August 7 and explicitly demonstrates cross-request score corruption from a reused request identifier. The detailed GitHub advisory followed September 23; CVE publication was October 5, with a substantive revision November 18. The current advisory identifies 0.30.0 as corrected while retaining its earlier 0.25.1 test narrative.

Sources