Executive brief
SGLang, an open-source framework for serving Large Language Models (LLMs), contains a critical security flaw in its optional debugging subsystem. If the 'dumper' feature is enabled, remote attackers can bypass security restrictions to execute arbitrary code on the server hosting the AI model. This could lead to a complete system takeover, theft of proprietary AI models, or unauthorized access to sensitive data processed by the inference engine.
Technical details
A remote code execution (RCE) vulnerability exists in SGLang's optional 'dumper' subsystem. The vulnerability is rooted in the use of an unsafe `eval()` call within a debugging component that lacks sufficient sandboxing. When the `DUMPER_SERVER_PORT` environment variable is set, the dumper HTTP control plane becomes active. An attacker can send specially crafted inference requests that trigger the dumper to evaluate malicious Python code, leading to a sandbox escape and execution of arbitrary commands with the privileges of the SGLang process. This is part of a broader architectural issue where administrative and debugging routes may lack enforced authentication if API keys are not explicitly configured.
Affected products
- SGLang Project SGLang <= v0.5.15
Timeline
- 2026-04-22: disclosed: Initial disclosure to maintainers via GitHub Security Advisories.
- 2026-07-30: advisory: Public disclosure and CERT/CC coordination.