The vLLM library contains a critical denial of service vulnerability that allows unauthenticated remote attackers to terminate decode workers by exploiting improper block count validation.

The vulnerability exists within the NIXL connector's prefix caching implementation, specifically in the NixlBaseConnectorWorker._apply_prefix_caching function. Due to a failure to validate block counts across multi-prompt completion requests, an unauthenticated attacker can trigger a reachable assertion failure (CWE-617) that results in the immediate termination of the decode worker process.

Successful exploitation of this vulnerability leads to a denial of service, rendering the affected vLLM decode worker unavailable until a manual or automated restart is performed. In high-availability AI inference environments, this disruption can lead to significant service degradation, loss of compute resources, and operational downtime. Given the CVSS score of 7.5, the vulnerability is classified as High severity due to its potential to impact system availability without requiring user interaction or authentication.

Immediate Action: Update the vLLM installation to a version beyond 0.29.0 as soon as the vendor provides a patch. If no patch is currently available, restrict access to the inference API to trusted internal networks to reduce the attack surface.

Proactive Monitoring: Monitor system logs for frequent worker process crashes or assertion errors related to NixlBaseConnectorWorker. Alert on repeated service restarts that coincide with high volumes of multi-prompt completion requests.

Compensating Controls: Deploy a Web Application Firewall (WAF) or API gateway to inspect and rate-limit incoming completion requests, specifically targeting requests with abnormally high prompt counts or complex multi-prompt structures that may trigger the flaw.

Public Exploit Available: No

This vulnerability presents a clear risk to the availability of inference services using the NIXL connector. Organizations should prioritize patching this component as soon as an updated version is released. In the interim, implement strict request validation and network-level access controls to mitigate the risk of remote disruption by unauthorized actors.