Overview

Two unsafe pickle deserialization vulnerabilities have been discovered in the SGLang open-source project, one within the tool's multimodal generation module and another within the Encoder Parallel Disaggregation system. SGLang is a serving framework for large language models (LLMs) and multimodal models. These vulnerabilities, tracked as CVE-2026-3059 and CVE-2026-3060, allow an attacker to provide a malicious pickle file to a device using SGLang's multimodal generation system or the encoder parallel disaggregation system, enabling code execution. CVE-2026-3989 allows an attacker to provide a malicious pickle file that, when attempted to be replayed with the replay_request_dump.py script, will result in code execution. Version 0.5.10 addresses the vulnerabilities. It is recommended that project maintainers avoid implementing Pickle functions due to the inherent security risks.

Description

SGLang is a framework for serving various large language models (LLMs) and multimodal AI models, supporting models such as Qwen, DeepSeek, Mistral, and Skywork, and is compatible with OpenAI APIs. Two unsafe pickle deserialization vulnerabilities have been identified in the project, tracked as CVE-2026-3059 and CVE-2026-3060.

CVE-2026-3059 SGLangs multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker, which deserializes untrusted data using pickle.loads() without authentication.

CVE-2026-3060 SGLangs encoder parallel disaggregation system is vulnerable to unauthenticated remote code execution through the disaggregation module, which deserializes untrusted data using pickle.loads() without authentication.

SGLang is vulnerable to CVE-2026-3059 when the multimodal generation system is enabled, and to CVE-2026-3060 when the encoder parallel disaggregation system is enabled. If either condition is met and an attacker knows the TCP port on which the ZMQ broker is listening and can send requests to the server, they can exploit the vulnerability by sending a malicious pickle file to the broker, which will then deserialize it.

CVE-2026-3989 SGLangs replay_request_dump.py contains an insecure pickle.load() without validation and proper deserialization. An attacker can take advantage of this by providing a malicious .pkl file, which will execute the attackers code on the device running the script.

The SGLang project's replay_request_dump.py script uses pickle.load() without trust validation, allowing for arbitrary code execution if an attacker can control the pickle file contents. This vulnerability has low applicability but high impact, and can be exploited if an attacker can provide a malicious pickle file or write to the crash dump directory, potentially through social engineering or by compromising a directory where crash dump information is automatically saved. The script, intended to replay crash dump information, poses a risk if an attacker can manipulate the input files or directories used by the script.

The use of Pickle is strongly discouraged due to its inherent security risks. Deserializing a pickle file with pickle.loads() can lead to Remote Code Execution (RCE) if an attacker can provide a malicious file. This is because the pickle file format stores not only data but also instructions on how to reconstruct the object, which are executed during deserialization. As a result, an attacker can potentially execute arbitrary Python code.

To mitigate these risks, projects should consider implementing safer serialization formats, such as JSON or XML, or using tools like msgpack to perform more data-driven serialization and deserialization instead of open-ended capabilities such as pickle. This can help prevent RCE attacks and ensure the secure exchange of data.

Impact

An attacker who can send crafted messages to the ZeroMQ interface may trigger unsafe pickle deserialization in SGLang when certain features are enabled, potentially leading to remote code execution. Additionally, an attacker who can provide a malicious pickle file in the context of a crash dump may also exploit CVE-2026-3989. Successful exploitation could allow arbitrary code execution in the context of the SGLang service, potentially leading to host compromise, lateral movement, data exfiltration, or denial-of-service (DoS) attacks. Deployments that expose the affected interface to untrusted networks are at the highest risk of exploitation.

Solution

Users of SGLang should restrict access to the service interfaces and ensure they are not exposed to untrusted networks. Proper network segmentation and access controls should be implemented to prevent unauthorized interaction with the ZeroMQ endpoints. Version 0.5.10 addresses the vulnerabilities.

Acknowledgements

Thanks to the reporter, Igor Stepansky. Thank you to Gregory Bowers for additional research assistance during coordination.This document was written by Christopher Cullen.

Vendor Information