In versions up to 0.29.0 of vLLM, the MooncakeConnector component creates temporary transfer placeholders when prefill requests are rejected. These placeholders are never removed, causing the sender task pool to become saturated. As a result, legitimate requests can be delayed by as much as 480 seconds, while health checks still indicate that the system is healthy. This behavior constitutes a resource exhaustion vulnerability that can degrade performance and availability. The weakness is consistent with the CWE-772 classification, reflecting improper control of data lifetime and resources.

vLLM version 0.29.0 and earlier that use the MooncakeConnector for distributed KV transfer. The product is distributed as the vllm package from the vllm-project.

The CVSS score of 6.9 indicates moderate severity, and the EPSS score is not available, so the exact likelihood of exploitation is uncertain. The vulnerability is publicly documented but not listed in the CISA KEV catalog, suggesting no confirmed large‑scale exploitation. Attackers can exploit the flaw by sending a volume of rejected prefill requests, thereby exhausting the task pool. Due to the lack of an official workaround, mitigation must rely on updating the software or disabling the vulnerable component.