A critical Pickle deserialization vulnerability exists in the SGLang project. This flaw allows attackers to execute arbitrary code remotely on vulnerable servers. Exploitation requires the expert-parallel backup subsystem to be enabled. Additionally, the attacker needs network access to the SGLang service. Currently, there is no patch available for this vulnerability. The project maintainers have not responded to coordination efforts. SGLang is an open-source framework for serving large language models and AI models. The expert-parallel backup subsystem uses a ZeroMQ PULL socket without authentication or deserialization safeguards. This allows attackers to send malicious pickle files leading to remote code execution. Users should restrict access to the service interfaces and avoid exposing them to untrusted networks. Setting SGLANG_USE_PICKLE_IPC to false in environ.py is a recommended mitigation.