CVE-2026-129356

vLLM could use another request’s cached prompt prefix

vLLM project · vLLM cascade attention and KV-cache offload

First documented disclosure · Review checkpoint

Summary

With cascade attention and CPU cache offload, vLLM could decode a request against another request’s cached context. The reporter observed content from the other context in the model’s output.

Severity and exploitation evidence

Low · 2.3/10 · CVSS 4.0. Severity source.

Exploitation not established: The reporter reproduced cross-request output on two releases in a controlled test. The reviewed sources identify no production victim.

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

Concurrent cache operations could cause cascade attention to treat unrelated requests as sharing a prefix.

Conditions required

Multiple requests shared a serving instance configured for cascade attention, prefix caching and the simple CPU offload connector. The reporter tested the FlashAttention backend.

Potential impact

A response could contain content from a different request or give an incorrect result without signaling an error.

Affected versions and configurations

0.30.0 and 0.31.0, reproduced upstream.

Fix and mitigation status

0.32.0. This release removes cascade attention.

Practical next steps

Upgrade vLLM to 0.32.0 or later. On a deployment awaiting upgrade, disable cascade attention.

Operating lesson

Test request isolation across concurrent cache offload and reuse.

Evidence and limits

The reporter opened the issue on October 14. Version 0.32.0 shipped on October 18. VulDB published the CVE on November 15 and rated it 2.3 under CVSS 4.0. EUVD lists it as EUVD-2026-125953.

Sources