Remote Code Execution in LightLLM via RPyC Deserialization

Executive Summary

This vulnerability is a remote code execution flaw in LightLLM through version 1.2.0. When the system is started with the --pd_trans_mode nccl option, it exposes an unauthenticated RPyC control channel in the KV-transfer worker. Attackers can send malicious pickled objects to this channel, which deserializes the data and executes arbitrary code with the privileges of the LightLLM service account.

Useful 0 Not Useful 0

Detection Guidance

Check if LightLLM is running with the --pd_trans_mode nccl parameter. Scan for open ports in the range 30000-40000 on prefill or decode nodes. Look for RPyC ThreadedServer processes listening on these ports. Verify if LightLLM version is 1.2.0 or earlier.

Useful 0 Not Useful 0

Impact Analysis

An attacker could gain full control over the LightLLM service, execute arbitrary commands, access sensitive data, or disrupt operations. Since the attack does not disrupt service, it may remain undetected while allowing the attacker to maintain persistence or exfiltrate data.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability could lead to unauthorized access to sensitive data, violating GDPR's data protection principles and HIPAA's security requirements. Organizations using affected LightLLM versions may face compliance violations, legal penalties, and reputational damage due to potential data breaches.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade LightLLM to a version beyond 1.2.0. Disable the --pd_trans_mode nccl parameter if used. Restrict network access to ports 30000-40000. Implement authentication for the RPyC control channel. Use safe serialization methods instead of pickle.

Useful 0 Not Useful 0