CVE•Published 2026-10-26•Modified 2026-10-28•0 articles on news•4 live references•NVD data

CVE-2026-115916Vllm · Vllm

Vulnerability data via NVD (ingested)

CVSS v3.1

6.5

MEDIUM

CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

EPSS percentile

—

Weaknesses (CWE)

CWE-129Improper Validation of Array Index CWE-193Off-by-one Error

Description

vLLM is an inference and serving engine for large language models. Prior to 0.32.0, vllm/tokenizers/hf.py computes max_token_id with an off-by-one upper bound, allowing a pre-tokenized prompt ID equal to the vocabulary size to pass validation. For a tensor-parallel size of one and an exact multiple-of-64 vocabulary, VocabParallelEmbedding.forward does not mask that value before F.embedding uses it, causing a CUDA device-side assertion and terminating the engine. This issue is fixed in version 0.32.0.

Timeline

Published 2026-10-26

Modified 2026-10-28

External references

NVD MITRE Sploitus VulnCheck

Search for exposed instances

Shodan + Censys queries derived from NVD's CPE data. The vuln tag catches assets Shodan has explicitly linked to this CVE; the product / banner fingerprints find exposed instances even when the vuln tag was never applied (which is common).

Shodan · vuln tag0 hosts

vuln:CVE-2026-115916

Hosts Shodan has explicitly fingerprinted as vulnerable.

Shodan · product

product:"Vllm Vllm"

All exposed Vllm Vllm instances — cross-reference with the CVE's affected-version range.

Shodan · banner/body mention

http.html:"Vllm"

HTTP body or banner mentions "Vllm" — catches deploys Shodan didn't identify as a product.

More intel sources (5)

Shodan report

vuln:CVE-2026-115916