Unauthenticated RCE Control in LightLLM via RL Endpoints

Executive Summary

LightLLM through version 1.2.0 exposes reinforcement learning control routes on its public HTTP API without authentication. Unauthenticated attackers can access endpoints like /pause_generation, /abort_request, /flush_cache, and /init_weights_update_group to disrupt inference operations and cause workers to become unresponsive.

Useful 0 Not Useful 0

Detection Guidance

Check if LightLLM's HTTP API exposes unauthenticated RL control routes by sending requests to endpoints like /pause_generation, /abort_request, /flush_cache, or /init_weights_update_group. Use curl commands such as: curl -v http://<target-ip>:<port>/pause_generation. If the server responds with a valid JSON or error (not 404), the vulnerability likely exists.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability allows attackers to disrupt model inference operations, causing service outages or making workers unresponsive. It can lead to denial-of-service conditions and potential integrity risks if weight-update endpoints are exploited.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade LightLLM to a patched version if available. If upgrading is not possible, restrict network access to the LightLLM HTTP port using firewalls or network policies. Disable the --enable_rl flag if RL features are not required. Monitor logs for suspicious requests to RL endpoints.

Useful 0 Not Useful 0