vLLM, an inference and serving engine for large language models, has a Token Injection vulnerability in its multimodal processing component in versions >= 0.6.1 and < 0.20.0. When unauthenticated text-only prompts include special tokens interpreted as control, and when image/video placeholder sequences lack corresponding data, the software attempts to index into empty grids during input-position computation. This improper validation of array indices (CWE-129) leads to an unhandled IndexError, causing worker termination or availability degradation. The vulnerability affects multimodal paths that use image_grid_thw or video_grid_thw. The issue is resolved in vLLM 0.20.0.

The vulnerability can cause denial of service by terminating worker processes or degrading availability of the vLLM service when malformed multimodal inputs are processed. There is no impact on confidentiality or integrity reported. No known exploits in the wild have been identified.

Upgrade vLLM to version 0.20.0 or later, where this vulnerability is fixed. Since the vendor advisory indicates the issue is resolved in 0.20.0, applying this official fix is the recommended remediation. No other mitigation steps are specified.