1. Home
  2. Vulnerabilities
  3. CVE-2026-54234
7.5
HIGH CVSS 3.1
CVE-2026-54234
vLLM: Remote DoS in vLLM via Invalid Recovered Token Reinjection
Overview
Description

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted to negative one when the engine selects the next live token for a request and is written back into the drafter's input ids; that out-of-vocabulary value is later consumed by the model's embedding and attention path and crashes the engine worker with a GPU device-side assertion. The same triggering request sequence is reachable through the public gRPC Generate and Abort endpoints, so a remote client that can send generation requests can crash the shared engine worker, aborting concurrent requests and causing a service-wide denial of service for other clients of the deployment until the worker is restarted. This issue is fixed in version 0.24.0.

Details
INFO

Published Date :

July 6, 2026, 9:16 p.m.

Last Modified :

July 6, 2026, 9:16 p.m.

Remotely Exploit :

Yes !

Source :

[email protected]
Impact
Affected Products

The following products are affected by CVE-2026-54234 vulnerability. Even if cvefeed.io is aware of the exact versions of the products that are affected, the information is not represented in the table below.

ID Vendor Product Action
1 Vllm-project vllm
: Total Affected Vendor : 1 | Products : 1
Scoring
CVSS Scores
The Common Vulnerability Scoring System is a standardized framework for assessing the severity of vulnerabilities in software and systems. We collect and displays CVSS scores from various sources for each CVE.
Score Version Severity Vector Exploitability Score Impact Score Source
CVSS 3.1 HIGH [email protected]
Solution
Update vLLM to version 0.24.0 to fix the speculative decoding denial of service vulnerability.
  • Update vLLM to version 0.24.0.
  • Apply the fix for speculative decoding.
References
References to Advisories, Solutions, and Tools

Here, you will find a curated list of external links that provide in-depth information, practical solutions, and valuable tools related to CVE-2026-54234.

URL Resource
https://github.com/vllm-project/vllm/commit/8a5cf1ccd65e8ac7635c402c1ec0b08988bc26ca
https://github.com/vllm-project/vllm/pull/44744
https://github.com/vllm-project/vllm/security/advisories/GHSA-8wr5-jm2h-8r4f
CWE - Common Weakness Enumeration

While CVE identifies specific instances of vulnerabilities, CWE categorizes the common flaws or weaknesses that can lead to vulnerabilities. CVE-2026-54234 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-2026-54234 weaknesses.

CAPEC-3: Using Leading 'Ghost' Character Sequences to Bypass Input Filters Using Leading 'Ghost' Character Sequences to Bypass Input Filters CAPEC-7: Blind SQL Injection Blind SQL Injection CAPEC-8: Buffer Overflow in an API Call Buffer Overflow in an API Call CAPEC-9: Buffer Overflow in Local Command-Line Utilities Buffer Overflow in Local Command-Line Utilities CAPEC-10: Buffer Overflow via Environment Variables Buffer Overflow via Environment Variables CAPEC-13: Subverting Environment Variable Values Subverting Environment Variable Values CAPEC-14: Client-side Injection-induced Buffer Overflow Client-side Injection-induced Buffer Overflow CAPEC-22: Exploiting Trust in Client Exploiting Trust in Client CAPEC-23: File Content Injection File Content Injection CAPEC-24: Filter Failure through Buffer Overflow Filter Failure through Buffer Overflow CAPEC-28: Fuzzing Fuzzing CAPEC-31: Accessing/Intercepting/Modifying HTTP Cookies Accessing/Intercepting/Modifying HTTP Cookies CAPEC-42: MIME Conversion MIME Conversion CAPEC-43: Exploiting Multiple Input Interpretation Layers Exploiting Multiple Input Interpretation Layers CAPEC-45: Buffer Overflow via Symbolic Links Buffer Overflow via Symbolic Links CAPEC-46: Overflow Variables and Tags Overflow Variables and Tags CAPEC-47: Buffer Overflow via Parameter Expansion Buffer Overflow via Parameter Expansion CAPEC-52: Embedding NULL Bytes Embedding NULL Bytes CAPEC-53: Postfix, Null Terminate, and Backslash Postfix, Null Terminate, and Backslash CAPEC-63: Cross-Site Scripting (XSS) Cross-Site Scripting (XSS) CAPEC-64: Using Slashes and URL Encoding Combined to Bypass Validation Logic Using Slashes and URL Encoding Combined to Bypass Validation Logic CAPEC-67: String Format Overflow in syslog() String Format Overflow in syslog() CAPEC-71: Using Unicode Encoding to Bypass Validation Logic Using Unicode Encoding to Bypass Validation Logic CAPEC-72: URL Encoding URL Encoding CAPEC-73: User-Controlled Filename User-Controlled Filename CAPEC-78: Using Escaped Slashes in Alternate Encoding Using Escaped Slashes in Alternate Encoding CAPEC-79: Using Slashes in Alternate Encoding Using Slashes in Alternate Encoding CAPEC-80: Using UTF-8 Encoding to Bypass Validation Logic Using UTF-8 Encoding to Bypass Validation Logic CAPEC-81: Web Server Logs Tampering Web Server Logs Tampering CAPEC-83: XPath Injection XPath Injection CAPEC-85: AJAX Footprinting AJAX Footprinting CAPEC-88: OS Command Injection OS Command Injection CAPEC-101: Server Side Include (SSI) Injection Server Side Include (SSI) Injection CAPEC-104: Cross Zone Scripting Cross Zone Scripting CAPEC-108: Command Line Execution through SQL Injection Command Line Execution through SQL Injection CAPEC-109: Object Relational Mapping Injection Object Relational Mapping Injection CAPEC-110: SQL Injection through SOAP Parameter Tampering SQL Injection through SOAP Parameter Tampering CAPEC-120: Double Encoding Double Encoding CAPEC-135: Format String Injection Format String Injection CAPEC-136: LDAP Injection LDAP Injection CAPEC-153: Input Data Manipulation Input Data Manipulation CAPEC-182: Flash Injection Flash Injection CAPEC-209: XSS Using MIME Type Mismatch XSS Using MIME Type Mismatch CAPEC-230: Serialized Data with Nested Payloads Serialized Data with Nested Payloads CAPEC-231: Oversized Serialized Data Payloads Oversized Serialized Data Payloads CAPEC-250: XML Injection XML Injection CAPEC-261: Fuzzing for garnering other adjacent user/sensitive data Fuzzing for garnering other adjacent user/sensitive data CAPEC-267: Leverage Alternate Encoding Leverage Alternate Encoding CAPEC-473: Signature Spoof Signature Spoof CAPEC-588: DOM-Based XSS DOM-Based XSS CAPEC-664: Server Side Request Forgery Server Side Request Forgery

We scan GitHub repositories to detect new proof-of-concept exploits. Following list is a collection of public exploits and proof-of-concepts, which have been published on GitHub (sorted by the most recently updated).

Results are limited to the first 15 repositories due to potential performance issues.

The following list is the news that have been mention CVE-2026-54234 vulnerability anywhere in the article.

Results are limited to the first 20 news articles due to potential performance issues.

The following table lists the changes that have been made to the CVE-2026-54234 vulnerability over time.

Vulnerability history details can be useful for understanding the evolution of a vulnerability, and for identifying the most recent changes that may impact the vulnerability's severity, exploitability, or other characteristics.

  • New CVE Received by [email protected]

    Jul. 06, 2026

    Action Type Old Value New Value
    Added Affected [{'vendor': 'vllm-project', 'product': 'vllm', 'versions': [{'status': 'affected', 'version': '< 0.24.0'}]}]
    Added Description vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted to negative one when the engine selects the next live token for a request and is written back into the drafter's input ids; that out-of-vocabulary value is later consumed by the model's embedding and attention path and crashes the engine worker with a GPU device-side assertion. The same triggering request sequence is reachable through the public gRPC Generate and Abort endpoints, so a remote client that can send generation requests can crash the shared engine worker, aborting concurrent requests and causing a service-wide denial of service for other clients of the deployment until the worker is restarted. This issue is fixed in version 0.24.0.
    Added CVSS V3.1 AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
    Added CWE CWE-20
    Added CWE CWE-1284
    Added Reference https://github.com/vllm-project/vllm/commit/8a5cf1ccd65e8ac7635c402c1ec0b08988bc26ca
    Added Reference https://github.com/vllm-project/vllm/pull/44744
    Added Reference https://github.com/vllm-project/vllm/security/advisories/GHSA-8wr5-jm2h-8r4f
EPSS is a daily estimate of the probability of exploitation activity being observed over the next 30 days. Following chart shows the EPSS score history of the vulnerability.
CVSS
Vulnerability Scoring Details
Base CVSS Score: 7.5
Attack Vector
Attack Complexity
Privileges Required
User Interaction
Scope
Confidentiality Impact
Integrity Impact
Availability Impact