LightLLM versions up to 1.2.0 allow unauthenticated callers to use the exposed_set_value API over the NCCL control channel to store key‑value pairs without any size limits, which drains worker process memory until the process crashes. The resulting crash triggers a node failure, disrupting availability of the inference service. This vulnerability is an instance of uncontrolled memory consumption.

The affected product is LightLLM from ModelTC, encompassing all releases through version 1.2.0. Users running this software with the --pd_trans_mode nccl option are susceptible if the NCCL control channel remains exposed to unauthenticated traffic.

The vulnerability carries a CVSS score of 8.7, indicating a high severity. EPSS is unavailable, and the issue is not listed in the CISA KEV catalog. The attack vector is remote and authenticated is not required, allowing any networked client with access to the NCCL control channel to trigger memory exhaustion. The primary impact is denial of service through worker process crashes, with no direct confidentiality or integrity compromise. Given the lack of mitigation controls in the exposed API, the exploitability is considered high for environments that expose the control channel to untrusted networks.