CVE-2026-93989

Received Received - Intake

vLLM SamplingParams Out-of-Bounds Token Index Vulnerability

Publication date: 2026-09-19

Last updated on: 2026-09-19

Assigner: VulnCheck

Description

vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in SamplingParams.update_from_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests, causing different in-flight HTTP requests to return incorrect tokens.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published

2026-09-19

Last Modified

2026-09-19

Generated

2026-09-20

AI Q&A

2026-09-20

EPSS Evaluated

N/A

NVD

CVE-2026-93989

Affected Vendors & Products

Currently, no data is known.

Helpful Resources

Exploitability

CWE

KEV

CWE ID	Description
CWE-129	The product uses untrusted input when calculating or using an array index, but the product does not validate or incorrectly validates the index to ensure the index references a valid position within the array.

Attack-Flow Graph

Executive Summary

vLLM through version 0.29.0 has a flaw in its SamplingParams.update_from_tokenizer() function where it fails to validate bad_words token indices against the model's output width. This allows attackers to provide out-of-bounds token indices that corrupt logits memory for concurrent requests, leading to incorrect tokens being returned in different in-flight HTTP requests.

Detection Guidance

Detection requires code review or runtime monitoring of vLLM instances. Check if SamplingParams.update_from_tokenizer() improperly validates bad_words token indices against model output width. Monitor for memory corruption or incorrect token returns in concurrent HTTP requests.

Impact Analysis

This vulnerability could cause applications using vLLM to return incorrect or corrupted responses to users. If exploited, it may lead to data processing errors, misinformation, or unexpected behavior in systems relying on vLLM for language model inference.

Mitigation Strategies

Upgrade vLLM to a patched version beyond 0.29.0. If upgrading is not possible, restrict concurrent HTTP requests and validate all token indices manually before processing. Monitor logs for memory corruption or incorrect token returns.

Hi! I’m here to help you understand CVE-2026-93989. Ask me anything about the vulnerability, its impact, or mitigation strategies.

0/70