Denial of Service Vulnerability in LightLLM

LightLLM versions through 1.2.0 contain a critical denial-of-service vulnerability that allows unauthenticated attackers to crash the inference worker. The vulnerability resides in the /generate endpoint, specifically within the handling of the internal pd_kv_trans_params field. By supplying a specially crafted payload in this parameter, an attacker triggers a failure in the pickle.loads function within the InferSamplingParams logic. This crash effectively terminates the inference thread. While the /health endpoint remains responsive, all subsequent legitimate requests to the inference service will hang, requiring a manual service restart to restore normal operations. This flaw represents a significant risk for deployments relying on LightLLM for inference tasks.

Impact

The vulnerability impacts any infrastructure running LightLLM versions up to and including 1.2.0. Successful exploitation results in a complete service outage for inference workloads. Because the /health check does not reflect the failure of the underlying inference thread, monitoring systems may fail to detect the outage immediately, potentially leading to prolonged periods of service unavailability for end users.

Recommendation

Patching is required to mitigate this risk. Ensure all LightLLM deployments are updated beyond version 1.2.0 to resolve the deserialization flaw. Until an update is applied, ensure that access to the /generate endpoint is restricted via network-level controls such as a Web Application Firewall or API gateway to prevent unauthenticated requests from reaching the vulnerable endpoint.