CVE-2026-94622: vLLM denial of service in NIXL connector metadata handling

Executive brief

vLLM is a high-performance inference engine used to serve large language models. In disaggregated deployments using the NIXL connector, an attacker can crash the decode engine by sending specially crafted requests with incomplete data, causing all pending requests to fail and requiring manual service restart. This impacts the availability of language model serving infrastructure.

Technical details

A KeyError is triggered in EngineCore scheduling when the kv_transfer_params dictionary contains incomplete entries during prefill/decode disaggregated processing. The vulnerability exists in the NIXL connector's metadata handling and can be exploited remotely by attackers crafting malformed requests. An attacker with network access to the service endpoint can terminate the decode engine, resulting in denial of service.

Affected products

• vLLM Project vLLM through 0.29.0

Timeline

• 2026-09-21: disclosed: CVE-2026-94622 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/nixl/metadata.py