CVE-2026-94627: vLLM Mooncake connector GPU KV cache memory leak

Executive brief

vLLM is an inference engine for large language models that manages GPU memory to cache key-value data during request processing. A flaw in the Mooncake connector's handling of concurrent requests causes GPU memory to leak, preventing new requests from being served until the service is restarted. An attacker can exhaust GPU memory by sending completion requests with multiple prompts, causing denial of service.

Technical details

The vulnerability exists in the Mooncake KV cache connector in vLLM through version 0.29.0, where concurrent child requests sharing a single transfer ID in disaggregated prefill/decode deployments fail to properly release GPU KV cache block ownership. This leads to orphaned cache blocks accumulating over time. An attacker can trigger GPU memory exhaustion by submitting concurrent completion requests, causing the process to become unresponsive until restart.

Affected products

• vLLM Project vLLM through 0.29.0

Timeline

• 2026-09-21: disclosed

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