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.
This vulnerability carries a HIGH severity rating with a CVSS v3.1 score of 7.5, indicating it can be exploited remotely over the network with relatively low complexity without requiring user interaction and does not require pre-existing privileges . The vulnerability impacts and availability (service disruption) for affected systems. Impacting 1 product from vllm organizations running these solutions should prioritize assessment and patching.
Reported in 2026, this vulnerability emerged during an era marked by increased sophistication in supply chain attacks, cloud infrastructure vulnerabilities, and software-as-a-service (SaaS) security challenges. Security practices during this period emphasized zero-trust architectures, container security, and API protection.
2026-07-06T21:16:56.477
2026-07-07T19:04:17.300
Analyzed
CVSSv3.1: 7.5 (HIGH)
SecUtils normalizes and enriches National Vulnerability Database (NVD) records by standardizing vendor and product identifiers, aggregating vulnerability metadata from both NVD and MITRE sources, and providing structured context for security teams. For vllm's affected products, we extract Common Platform Enumeration (CPE) data, Common Weakness Enumeration (CWE) classifications, CVSS severity metrics, and reference data to enable rapid vulnerability prioritization and asset correlation. This record contains no exploit code, proof-of-concept instructions, or attack methodologies—only defensive intelligence necessary for patch management, risk assessment, and security operations.