CVE-2026-103243

LightLLM through version 1.2.0 mounts reinforcement learning control routes on the public HTTP API without authentication checks. Unauthenticated attackers can call endpoints like /pause_generation, /abort_request, /flush_cache, and /init_weights_update_group to disrupt inference operations and wedge workers on deployments started with --enable_rl.

LightLLM through 1.2.0 visual_only deployments expose an unauthenticated RPyC service with allow_pickle enabled that deserializes attacker-supplied arguments in the remote_infer_images method. Attackers can reach the visual RPyC port and pass objects with __reduce__ methods to execute arbitrary code with service account privileges.

LightLLM through 1.2.0 multimodal deployments expose an unauthenticated RPyC cache service with pickle deserialization enabled on all interfaces. Attackers can send crafted serialized objects to exposed cache methods to execute arbitrary code with service privileges.

LightLLM through 1.2.0 contains a remote code execution vulnerability in the router profiler service when started with --enable_profiling flag. The service exposes an unauthenticated RPyC server with pickle deserialization enabled, allowing attackers to execute arbitrary code by sending crafted serialized objects to the profiler command queue.

LightLLM through 1.2.0 contains a remote code execution vulnerability in the KV-transfer worker when started with --pd_trans_mode nccl, which exposes an unauthenticated RPyC control channel that deserializes attacker-supplied data. Attackers can send malicious pickled objects to the exposed RPyC ThreadedServer to execute arbitrary code with the privileges of the LightLLM service account.

LightLLM through version 1.2.0 contains an authentication bypass vulnerability in the /pd_register WebSocket endpoint that allows unauthenticated attackers to register arbitrary nodes by supplying crafted JSON without peer address validation. Attackers can disclose full user prompts routed to their socket, trigger denial of service by replacing legitimate nodes, or make the PD Master issue requests to internal network addresses.

LightLLM through 1.2.0 contains a remote code execution vulnerability in the Config Server's unauthenticated /visual_register WebSocket endpoint, which passes the first client frame directly to pickle.loads(). Attackers can reach the Config Server port and send a malicious serialized payload with a __reduce__ method to execute arbitrary code with Config Server process privileges.

LightLLM version 1.1.0 and prior contain an unauthenticated remote code execution vulnerability in PD (prefill-decode) disaggregation mode. The PD master node exposes WebSocket endpoints that receive binary frames and pass the data directly to pickle.loads() without authentication or validation. A remote attacker who can reach the PD master can send a crafted payload to achieve arbitrary code execution.

LightLLM through 1.2.0 contains a denial of service vulnerability. An unauthenticated attacker can send a single request to the /generate endpoint with the internal pd_kv_trans_params field, causing pickle.loads to fail in InferSamplingParams and killing the inference thread while /health stays up and later requests hang until restart.

LightLLM through version 1.2.0 contains a memory exhaustion vulnerability in the NCCL control channel when started with --pd_trans_mode nccl. Unauthenticated attackers can call the exposed_set_value method to store unbounded key-value pairs without size limits, causing the worker process to crash and triggering node failure.