SGLang, a framework for serving large language models, contains a remote code execution vulnerability due to unsafe deserialization of pickle data in its expert-parallel backup subsystem. The subsystem exposes a ZeroMQ PULL socket on a routable network interface without authentication, allowing any network-accessible attacker to send malicious pickle payloads that get deserialized, leading to unauthenticated remote code execution. This vulnerability is tracked as CVE-2026-14890 and is similar in nature to CVE-2026-7301 and CVE-2026-7304 but affects a different subsystem. The vulnerability requires that the expert-parallel backup subsystem be enabled and accessible over the network. No official patch is available, and the maintainers are working on refactoring the codebase to use msgpack instead of pickle for safer deserialization. Until a patch is released, users should disable the pickle IPC feature and restrict network access to the service.

An unauthenticated attacker with network access to the expert-parallel backup subsystem of SGLang can achieve remote code execution on the host running the vulnerable service. Deployments exposing this interface to untrusted networks are at high risk. This could lead to full compromise of the affected system.

No official patch or fix is currently available. Users should immediately restrict network access to the expert-parallel backup subsystem to trusted networks only and implement network segmentation and access controls to prevent unauthorized access. Additionally, users should set the environment variable SGLANG_USE_PICKLE_IPC to "false" in environ.py to disable the vulnerable pickle deserialization feature. Monitor the SGLang project for updates as maintainers are working to replace pickle with safer serialization methods such as msgpack.