rootpwn

high · CVSS v3 5.9 · CVSS v4 8.2

CVE-2026-100652

vLLM versions 0.22.0 through 0.23.0 have a validation flaw that allows out-of-vocabulary stop_token_ids to be processed, causing CUDA tensor

Overview

vLLM versions 0.22.0 through 0.23.0 have a validation flaw that allows out-of-vocabulary stop_token_ids to be processed, causing CUDA tensor indexing failures. This can crash the EngineCore, requiring a service restart. The issue affects deployments using the Rust HTTP or gRPC frontends.

Description

vLLM versions 0.22.0 through 0.23.0 fail to validate stop_token_ids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor. Attackers can submit requests with min_tokens greater than zero and out-of-vocabulary stop_token_ids to trigger CUDA tensor indexing failures that leave EngineCore in a fatal state requiring service restart.

Impact

Confidentiality: no direct data exposure. Integrity: potential denial of service by crashing EngineCore. Availability: service interruption until restart. Impacted parties: organizations running vLLM inference services, especially those using Rust frontends.

Remediation

Upgrade to vLLM 0.23.1 or later where stop_token_ids bounds are validated. If upgrade not possible, disable or sanitize stop_token_ids in requests, enforce min_tokens=0, or implement request validation to reject out-of-range IDs. Monitor logs for CUDA tensor errors and restart services automatically.

Risk context

High severity with CVSS v4 score of 8.2 indicates a significant risk. No EPSS data available, but the flaw can cause service crashes, so timely patching is recommended.

Affected products

  • vLLM 0.22.0
  • vLLM 0.23.0
  • Rust HTTP frontend
  • Rust gRPC frontend

Scores

Severity
high
CVSS v2
5.4
CVSS v3
5.9
CVSS v4
8.2
EPSS
—

vLLM Rust CUDA Denial-of-Service TensorIndexing Frontends Patch ServiceCrash

← All CVEs