LightLLM versions up to 1.2.0 allow an unauthenticated attacker to send a crafted request to the /generate endpoint that causes the internal pd_kv_trans_params field to be deserialized with pickle.loads. The deserialization failure terminates the inference worker thread while the health check endpoint remains healthy, leading to hung requests and requiring a restart before the service resumes normal operation. This flaw results in a loss of availability for the inference service without compromising data confidentiality or integrity.

The vulnerability affects ModelTC's LightLLM product for all releases up to and including version 1.2.0. Systems running these versions without a later, patched release are susceptible when the /generate API is exposed to external traffic.

With a CVSS score of 8.7 the issue is high severity, and the EPSS score below 1% indicates a low probability of widespread exploitation at present. The flaw is network‑accessible and requires no authentication, making it straightforward to trigger from any host that can reach the inference API. Although the vulnerability is not listed in CISA’s KEV catalog, its impact on service availability and the ease of exploitation warrant prompt attention.