LightLLM releases up to version 1.2.0 expose reinforcement‑learning (RL) control routes on the public HTTP API without any authentication. The exposed endpoints—such as /pause_generation, /abort_request, /flush_cache and /init_weights_update_group—can be invoked by anyone who reaches the server, allowing an attacker to interrupt or sabotage inference workloads. This flaw leads to operational disruption rather than data exfiltration, as the attacker can stop or reset workers that are serving user requests.

The vulnerability affects installations of LightLLM up through version 1.2.0 provided by ModelTC. No further version filtering is listed, indicating that the entire public API surface in those releases is vulnerable.

The CVSS score of 8.7 reflects a high‑severity risk, and though an EPSS score is not published, the flaw remains unlisted in the CISA KEV catalog. The likely attack vector is a network‑based request to the public HTTP API, so any machine with network reach to the server could exploit it. Once invoked, the attacker gains full control over the RL control plane, causing a denial of service to legitimate inference traffic.