CVE-2026-54234 vllm-project CVE debrief

Who should care

Users of vLLM versions prior to 0.24.0 should be aware of this vulnerability and take steps to mitigate it. This includes updating to version 0.24.0 or later and implementing compensating controls to detect and prevent similar attacks. Operators, platform administrators, vulnerability management teams, and security teams should review the vulnerability's impact and take necessary actions.

Technical summary

The vulnerability exists in the vLLM engine's handling of speculative decoding workloads. A remote client can send a specific sequence of requests to the engine, causing it to crash and resulting in a denial of service for other clients. The issue is fixed in version 0.24.0. The engine's embedding and attention path crash when consuming the out-of-vocabulary value produced by the rejection sampler.

Defensive priority

High

Recommended defensive actions

• Update vLLM to version 0.24.0 or later
• Implement compensating controls to detect and prevent similar attacks
• Monitor the engine for suspicious activity
• Restrict access to the engine to authorized clients only
• Review the supplied official advisory or CVE record to validate affected scope, severity, and vendor guidance
• Confirm whether affected product deployments exist in managed environments and assign an owner for follow-up
• Track exceptions, retest remediated assets, and close the item only after evidence is documented

Evidence notes

The CVE record was published on 2026-07-06T21:16:56.477Z and has been modified since then. The NVD entry is currently Analyzed. The vulnerability has a CVSS score of 7.5 and a severity of HIGH. Evidence is limited, and defenders should verify the vulnerability's impact and affected scope. The rejection sampler can 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.

Official resources