Off-by-One Vulnerability in vLLM by vllm-project
CVE-2026-115916

6.5MEDIUM

Key Information:

Vendor

Vllm-project

Status
Vllm
Vendor
CVE Published:
26 October 2026

What is CVE-2026-115916?

The vLLM inference and serving engine for large language models contains an off-by-one vulnerability in its handling of tokenized prompts. Specifically, prior to version 0.32.0, the implementation computes the maximum token ID with an incorrectly set upper bound. This flaw allows a pre-tokenized prompt ID that matches the vocabulary size to bypass validation checks. In scenarios involving a tensor-parallel size of one and an exact multiple-of-64 vocabulary, this issue leads to VocabParallelEmbedding.forward failing to properly mask this value before it is utilized by F.embedding. Consequently, this can trigger a CUDA device-side assertion, resulting in the termination of the engine. This vulnerability has been addressed in version 0.32.0.

Affected Version(s)

vllm < 0.32.0

References

https://github.com/vllm-project/vllm/security/advisories/...
x_refsource_CONFIRM
https://github.com/vllm-project/vllm/pull/59491
x_refsource_MISC
https://github.com/vllm-project/vllm/commit/37cecc8141e2d...
x_refsource_MISC

CVSS V3.1

Score:
6.5
Severity:
MEDIUM
Confidentiality:
None
Integrity:
None
Availability:
None
Attack Vector:
Network
Attack Complexity:
Low
Privileges Required:
Low
User Interaction:
None
Scope:
Unchanged

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.