CVE-2026-94624: vLLM denial of service in P2P KV offloading

Executive brief

vLLM is an inference engine for large language models that optimizes serving performance. When configured for peer-to-peer KV offloading with tiering, attackers can trigger a denial of service by supplying arbitrary remote host and port values, causing the inference service to crash and stop processing all requests.

Technical details

A denial of service vulnerability exists in vLLM's P2P KV offloading mechanism when OffloadingConnector is configured with TieringOffloadingSpec. Attackers can inject malicious kv_transfer_params containing unreachable peer addresses, which creates ZeroMQ socket connections that are never cleaned up, eventually exhausting the context quota and triggering an unhandled ZMQError that crashes EngineCore. This requires the affected tiering configuration to be enabled but no authentication is needed.

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/v1/kv_offload/tiering/p2p/control/zmq.py