CVE-2026-114051

vLLM generation can stop while its health probe stays green

vLLM · Token-in/token-out generate endpoint

First documented disclosure · Review checkpoint

Summary

An invalid minimum token count can disable generation while the process continues answering health checks.

Severity and exploitation evidence

High · 7.1/10 · CVSS 4.0. Severity source.

Exploitation not established: The researchers describe a controlled hang. Reviewed evidence does not show malicious exploitation.

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 API defers validation when max_tokens is null. The engine fills the default and revalidates later, where an uncaught error kills its input thread.

Conditions required

Network access to /inference/v1/generate when mounted. A deployment requiring an API key or running an authenticating proxy requires authorized inference access.

Potential impact

Subsequent generation requests hang until the service restarts. The standard OpenAI routes reject this invalid combination before the shared engine.

Affected versions and configurations

Versions before 0.31.0 per VulnCheck and vLLM. The researcher identified the route in 0.29.0 and in configurations that enable it.

Fix and mitigation status

0.31.0 per vLLM.

Practical next steps

Upgrade and monitor a small end-to-end generation request alongside the HTTP health probe. Restrict direct access to the scale-out endpoint.

Operating lesson

Validate dependent limits after defaults are applied and probe the work path that clients actually need.

Evidence and limits

The public September 19 fix predates the October 6 advisory. Deployment access controls affect whether an API key is required.

Sources