CVE-2026-94625: vLLM resource exhaustion in MooncakeConnector

Executive brief

vLLM is a high-throughput inference engine for large language models that serves AI applications. A resource exhaustion vulnerability in the MooncakeConnector component allows attackers to send rejected requests that create orphaned placeholders, depleting sender task pools and delaying legitimate requests by up to 480 seconds while the service appears healthy. This can degrade service availability and cause request timeouts for valid users.

Technical details

The MooncakeConnector fails to properly reclaim transfer placeholders when prefill requests are rejected, leading to unbounded resource accumulation. An attacker can exploit this via network by sending malformed or rejected prefill requests, causing the sender task pool to become exhausted. The vulnerability affects vLLM through version 0.29.0 and results in denial of service conditions while health checks continue to return success, masking the degradation.

Affected products

• vLLM Project vLLM through 0.29.0

Timeline

• 2026-09-21: disclosed: CVE-2026-94625 published

References

• https://github.com/vllm-project/vllm
• https://github.com/vllm-project/vllm/blob/v0.29.0/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py