Six critical vulnerabilities have been identified in the SGLang project, an open-source framework for serving AI models. These vulnerabilities include unauthenticated remote code execution, server-side request forgery, and local file reading capabilities. Additionally, credential leakage and model weight exfiltration are possible through specific endpoints. Most of these vulnerabilities do not require authentication to exploit. At the time of this report, no patches have been released by the SGLang maintainers, and attempts to coordinate have been unsuccessful. One RCE vulnerability arises from bypassing SafeUnpickler's denylist, allowing arbitrary command execution. Another RCE is present when the dumper subsystem is enabled, permitting sandbox escapes. The multimodal generation endpoint is susceptible to SSRF and local file reads due to unsanitized image URL inputs. Loading model weights from HuggingFace repositories also presents an RCE risk via pickle deserialization. The server_info endpoint can leak API keys and SSL key information. Finally, model weights can be exfiltrated if no API keys are configured. Until patches are available, users are advised to restrict network access, implement network segmentation, and disable unused endpoints.