A critical remote code execution vulnerability in SGLang allows unauthenticated attackers to escape the sandbox and execute arbitrary code on systems running the dumper subsystem.

This vulnerability involves an improper neutralization of directives in dynamically evaluated code (CWE-95), which occurs when the optional dumper subsystem is enabled. An unauthenticated attacker can trigger this flaw during inference requests to achieve code execution.

The ability for an unauthenticated remote attacker to execute arbitrary code poses a catastrophic risk to confidentiality, integrity, and availability. With a CVSS score of 9.8, this flaw could lead to total system compromise, unauthorized access to sensitive inference data, and potential lateral movement within the network.

Immediate Action: Disable the optional dumper subsystem by ensuring the DUMPER_SERVER_PORT is not configured or exposed in your environment until an official patch is released.

Proactive Monitoring: Review system logs for unusual inference requests or unexpected process execution patterns originating from the dumper subsystem configuration.

Compensating Controls: Implement strict network segmentation to ensure that SGLang instances are not exposed to the public internet and restrict access to the relevant ports via host-based firewalls.

Public Exploit Available: Unknown.

Given the critical severity and the potential for complete system takeover, organizations must prioritize the immediate modification of their SGLang configuration to disable the vulnerable dumper subsystem. Security teams should maintain high vigilance and monitor vendor security advisories for the release of an official patch to remediate the underlying code injection flaw.