CVE-2026-54234 describes an improper input validation vulnerability in vLLM before version 0.24.0. A frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a token equal to the vocabulary size boundary, which is converted to -1 and written back into the drafter's input IDs. This out-of-vocabulary token is then processed by the model's embedding and attention mechanisms, causing a GPU device-side assertion failure that crashes the engine worker. Because the triggering request sequence is accessible via the public gRPC Generate and Abort endpoints, a remote unauthenticated attacker can cause a denial of service by crashing the shared engine worker, disrupting all concurrent requests until the worker is restarted. The vulnerability is resolved in vLLM version 0.24.0.

Exploitation of this vulnerability results in a denial of service condition affecting the vLLM inference engine. A remote attacker can crash the engine worker by sending crafted generation requests, causing service-wide disruption until the worker is manually restarted. There is no impact on confidentiality or integrity, but availability is severely affected.

Upgrade vLLM to version 0.24.0 or later, where this vulnerability is fixed. There is no official patch or temporary fix other than upgrading. Until upgraded, restrict access to the public gRPC Generate and Abort endpoints to trusted clients to reduce exposure.