CVE-2026-44222

The record

vLLM is an inference and serving engine for large language models (LLMs). From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM’s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. This vulnerability is fixed in 0.20.0.

Affected packages

Related vulnerabilities

• CVE-2026-96675 alsa-lib through 1.2.16.1 contains a denial of service vulnerability in the multi PCM… CWE-129
• CVE-2026-84445 gRPC-Go is the Go language implementation of gRPC. CWE-129
• CVE-2026-88052 Tesseract is an open source OCR engine. CWE-129
• CVE-2026-87500 Improper validation of array index in ANGLE in Google Chrome prior to 153.0.8010.36… CWE-129
• CVE-2026-0799 In BPF instructions that load/store a value from/to a scratch memory register the… CWE-129
• CVE-2026-57159 PJSIP is a free and open source multimedia communication library written in C. CWE-129
• CVE-2026-85084 Out-of-bounds Write and Improper Validation of Array Index vulnerability in Samsung Open… CWE-129
• CVE-2023-31308 A malicious virtual function can invoke the certain command handlers in the SMU, causing… CWE-129