Executive brief
vLLM is an engine used to serve Large Language Models (LLMs) to users and applications. A flaw in how the system validates mathematical parameters allows specially crafted inputs (such as "Not a Number" or "Infinity") to bypass security checks. If exploited, these values can cause the system's processing units (GPUs) to crash, leading to a service outage for all users.
Technical details
A vulnerability exists in vLLM's input validation logic within `sampling_params.py`. The validation gates for 'temperature' and 'repetition_penalty' parameters rely on standard Python comparison operators (<, >), which return False for NaN and positive Infinity values according to IEEE 754 float semantics. Because these comparisons fail silently without triggering existing validation errors, the non-finite values propagate to GPU sampling kernels. This results in undefined behavior or CUDA errors that crash the inference worker process. The issue is resolved in version 0.23.1rc0 by implementing `math.isfinite()` checks.
Affected products
- vLLM Project vLLM < 0.23.1rc0
Timeline
- 2026-06-10: other: Fix proposed in pull request
- 2026-06-11: advisory: GitHub Security Advisory published
- 2026-06-22: disclosed: CVE published to NVD