CVE-2026-54234: high-severity vulnerability in vllm-project vllm
vLLM: Remote DoS in vLLM via Invalid Recovered Token Reinjection
Published · Updated
E-mail alert when it lands in KEV, gets a public PoC or its score goes up.
21Vexday Risk Score
No sign of exploitation. No public exploitation artifact known so far.
ssvc Trackcvss 7.5epss 0.6%
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.
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
Affected products
vllm-project · vllmRelated CVEs — vllm-project vllm
In the same product, most dangerous first.
CVE-2026-22778CRITICALvLLM leaks a heap address when PIL throws an errorEPSS 18.3%CVE-2025-32444CRITICALvLLM Vulnerable to Remote Code Execution via Mooncake IntegrationEPSS 1.8%CVE-2026-27893HIGHvLLM's hardcoded trust_remote_code=True in NemotronVL and KimiK25 bypasses user security opt-outEPSS 1.4%CVE-2026-48746CRITICALvLLM: OpenAI auth bypassEPSS 1.2%CVE-2026-22807HIGHvLLM affected by RCE via auto_map dynamic module loading during model initializationEPSS 1.0%CVE-2025-47277CRITICALvLLM Allows Remote Code Execution via PyNcclPipe Communication ServiceEPSS 1.0%