Back to the CVE list
CVE-2026-115916

vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)

CVSS score

6.5

EPSS probability

—

Risk score

0.7

Published

2 h ago

Patch status has not yet been verified

Resmî kaynaklarda yeterli kanıt bulunamadı. Bu sonuç “yama yok” anlamına gelmez.

Impact on My Inventory

Only the personal products you have added to your account are assessed; an organisation's inventory does not appear here.

Checking session…

Vulnerability description

DayBreach AI Explanation

Verified against sources

vLLM Denial of Service Vulnerability Prior to Version 0.32.0

Sources read

1

Source agreement

Tam uyum

Evidence confidence

%55

Short explanation

An input validation flaw in vLLM allows an authenticated attacker to trigger a CUDA device-side assertion, terminating the engine and causing a denial of service.

Why does it matter?

This vulnerability directly impacts system availability by causing the large language model serving engine to crash and resulting in a service outage.

What should be done?

  • Upgrade to version 0.32.0 or later
  • Follow vendor security advisories
Chain of evidence and sources3 verified claims

vLLM 0.32.0 öncesi sürümlerde, off-by-one hatası sözlük boyutundaki belirteç kimliklerinin doğrulamayı geçmesine izin verir.

VocabParallelEmbedding.forward işlevi değeri maskelemez ve F.embedding kullanımı sırasında CUDA cihaz tarafı iddiasını tetikler.

Bu güvenlik açığı motoru sonlandırarak hizmet engelleme durumuna yol açar ve 0.32.0 sürümünde giderilmiştir.

Show the technical text from the sources

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.

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

CWE: CWE-129, CWE-193

This explanation was compiled automatically from the source records. Scores, active exploitation, patch and exploit status come straight from the sources.

IMPACT: Affected Products and Software

Affected Products

The following products are affected by CVE-2026-115916 vulnerability. Where our threat engine knows the exact affected versions, they are listed below.

IDVENDORPRODUCTACTION
1vllm-projectvllm View
Total affected vendors: 1|Products: 1

RED HAT CSAF/VEX: Product Impact Check

Red Hat has not yet published a CSAF/VEX impact record for this CVE. This does not mean “not affected”.

SCORING: CVSS from Multiple Sources

CVSS Scores

The Common Vulnerability Scoring System is a standardised framework for assessing the severity of vulnerabilities in software and systems. We collect and display CVSS scores from several sources for each CVE.

SCOREVERSIONSEVERITYVECTORSOURCE
6.5CVSS V3_1MEDIUMCVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:HCVE CNA
6.5CVSS V31MEDIUMCVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:HNVD

Patch and Remediation Guide

Solution & Remediation Advisory

Patch status unknown

İncelenen resmî kaynaklarda yama durumu doğrulanamadı. Bu sonuç “yama yok” anlamına gelmez.

Confirm the affected product and version against the source evidence above.
Restrict access to internet-facing services and apply network segmentation.
Until a patch is confirmed, follow the vendor advisory; do not read an unknown status as “no patch”.
Monitor server logs and network endpoints for signs of anomalous execution.

Public exploit and active exploitation status

Verified exploit intelligence

No verified public PoC, exploit repository or Metasploit module has been found for this CVE yet.

Weakness Classification and Attack Patterns (CWE & CAPEC)

CWE - Common Weakness Enumeration

While CVE identifies specific instances of vulnerabilities, CWE categorises the common flaws or weaknesses that can lead to vulnerabilities. CVE-2026-115916 is associated with the following CWEs:

Common Attack Pattern Enumeration and Classification (CAPEC)

Common Attack Pattern Enumeration and Classification (CAPEC) stores attack patterns, which are descriptions of the common attributes and approaches employed by adversaries to exploit the CVE-2026-115916 weaknesses.

Attack Vector and Accessibility

ExposureAğ üzerinden
PermissionsLow
UserGerekmiyor

Official Sources and Advisories

CVSS Vector Radar Chart

6.5 / 10
Attack VectorNetworkComplexityLowPrivilegesLowUser Interact.NoneConfidentialityLowIntegrityNoneAvailabilityHigh

Vulnerability Scoring Details

CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Attack VectorNetwork
Attack ComplexityLow
Privileges RequiredLow privileges
User InteractionNone required
Confidentiality (Gizlilik)Low
Integrity (Bütünlük)None
Availability (Erişilebilirlik)High

Yama durumu bilinmiyor

İncelenen resmî kaynaklarda yama durumu doğrulanamadı. Bu sonuç “yama yok” anlamına gelmez.

Follow the vendor

Weakness Type (CWE & CAPEC)

Similar Vulnerabilities

Get this vulnerability through the API

Affected products, version ranges, exploitation status and patch information in a single request, with a free key.

curl -H "x-api-key: $DAYBREACH_API_KEY" \
  "https://api.daybreach.dev/api/v1/intel/cves/CVE-2026-115916"
CVE-2026-115916 · vLLM: Out-of-vocabulary prompt token id…