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

The vulnerability allows a remote unauthenticated attacker to cause a denial of service by crashing the engine worker process. This results in service interruption for all clients using the affected vLLM deployment until the worker is manually or automatically restarted. There is no impact on confidentiality or integrity reported.

Upgrade vLLM to version 0.24.0 or later, where this vulnerability is fixed. No other mitigation or workaround is documented. Until upgrading, restrict access to the gRPC Generate and Abort endpoints to trusted clients only to reduce exposure.