CVE-2026-90713

Deferred Deferred - Pending Action

Denial of Service in vLLM up to 0.29.0

Publication date: 2026-09-14

Last updated on: 2026-09-15

Assigner: VulDB

Description

A security flaw has been discovered in vllm-project vLLM up to 0.29.0. The affected element is the function TiktokenTokenizer::new of the file rust/src/text/src/backend/hf/mod.rs of the component tiktoken vocab File Handler. The manipulation results in denial of service. The attack is only possible with local access. The exploit has been released to the public and may be used for attacks. The pull request to fix this issue awaits acceptance.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published

2026-09-14

Last Modified

2026-09-15

Generated

2026-09-15

AI Q&A

2026-09-14

EPSS Evaluated

N/A

NVD

CVE-2026-90713

EUVD

EUVD-2026-77463

Affected Vendors & Products

Vendor	Product	Version / Range
vllm-project	vllm	to 0.29.0 (inc)
vllm-project	vllm-tokenizer	*
vllm-project	vllm-tokenizer	to 0.9.1 (inc)

Helpful Resources

Exploitability

CWE

KEV

CWE ID	Description
CWE-404	The product does not release or incorrectly releases a resource before it is made available for re-use.

Attack-Flow Graph

Executive Summary

This vulnerability is a denial of service flaw in the vLLM project up to version 0.29.0. It occurs in the TiktokenTokenizer::new function when processing a Tiktoken vocab file that assigns the same rank to two different tokens. Instead of returning an error, the function panics and crashes the system. The issue stems from a Rust panic in the tiktoken-rs library when encoder and decoder maps have unequal lengths due to duplicate ranks.

Detection Guidance

To detect this vulnerability, inspect vLLM installations for the affected TiktokenTokenizer::new function in rust/src/text/src/backend/hf/mod.rs. Check for malformed vocab files with duplicate token ranks using Rust-based tools or custom scripts that parse vocab files for rank conflicts.

Impact Analysis

This vulnerability allows an attacker with local access to crash the application by providing a malformed Tiktoken vocab file with duplicate token ranks. The crash results in a denial of service, making the system unavailable until restarted. It does not lead to data breaches or privilege escalation but disrupts service.

Compliance Impact

This vulnerability primarily impacts availability, which is a key aspect of compliance for standards like GDPR and HIPAA. A denial of service could lead to service unavailability, potentially violating availability requirements. However, there is no evidence of data exposure or confidentiality breaches from this specific flaw.

Mitigation Strategies

Apply the pending pull request fix (PR #51135) which adds validation to reject duplicate ranks in vocab files. Until then, avoid using untrusted vocab files and monitor for crashes in vLLM processes triggered by malformed input.

Hi! I’m here to help you understand CVE-2026-90713. Ask me anything about the vulnerability, its impact, or mitigation strategies.

0/70