CVE-2026-15974 – SGLang

This update addresses multiple vulnerabilities affecting SGLang. CVE-2026-15969 has a CVSS score of 9.8, Critical severity, and allows unauthenticated remote code execution through crafted pickle payloads that bypass SafeUnpickler protections. CVE-2026-15971 has a CVSS score of 9.8, Critical severity, with a verified public proof of concept (PoC) demonstrating remote code execution when the optional dumper subsystem is enabled. CVE-2026-15974 has a CVSS score of 6.5, Medium severity, with a verified PoC for server-side request forgery (SSRF) and local file disclosure through unsanitized image_url input. CVE-2026-15976 has a CVSS score of 9.8, Critical severity, with a verified PoC showing remote code execution during model weight loading from a Hugging Face repository. CVE-2026-15977 has a CVSS score of 7.5, High severity, with a verified PoC demonstrating exposure of API keys and SSL keyfile information through the /server_info endpoint. CVE-2026-15978 has a CVSS score of 7.5, High severity, with a verified PoC showing that attackers can exfiltrate model weights when API keys are not configured.

The update resolves vulnerabilities affecting model loading, deserialization, API endpoint security, credential protection, and distributed model handling. Successful exploitation could lead to remote code execution, credential disclosure, server-side request forgery, local file access, and theft of proprietary AI model weights. Several of the vulnerabilities require specific features or configurations to be enabled, but together they represent a significant risk to AI inference environments. Organizations should update to the latest supported version of SGLang as soon as possible.