CVE-2026-44222
From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM’s multimodal processing.
Does this matter?
High impact if exploited, but EPSS currently rates exploitation as unlikely (0.41%). Schedule it in the normal patch cycle and watch for a rise in EPSS or a public exploit.
Description
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.
- CVSS 3.1
- 7.5 HIGHCVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
- EPSS
- 0.41% probability · 35th percentile
- CISA KEV
- Not listed
- Weakness
- CWE-129
- Affected
- vllm/vllm
- Source
- security-advisories@github.com
References
- https://github.com/vllm-project/vllm/issues/32656Issue Tracking
- https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59fExploit, Vendor Advisory
Source: NVD record, EPSS from FIRST.org, KEV from CISA. Refreshed daily.