SGLang's multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker, which deserializes untrusted data using pickle.loads() without authentication.

SGLang' 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's reranking endpoint (/v1/rerank) achieves Remote Code Execution (RCE) when a model file containing a malcious tokenizer.chattemplate is loaded, as the Jinja2 chat templates are rendered using an unsandboxed jinja2.Environment().

SGLangs multimodal generation runtime is vulnerable to unauthenticated remote code execution when the --enable-custom-logit-processor option is enabled, as Python objects loaded via dill.loads() will be deserialized without validation.

SGLangs multimodal generation runtime scheduler's ROUTER socket binds to 0.0.0.0 by default and contains a sink that calls pickle.loads() on incoming messages, enabling RCE when exposed to the internet.

SGLang contains a RCE vulnerability when attempting to load model weights from a HuggingFace repository, specifically within the /updateweightsfromdisk, where torch.load(..., weightsonly=False) fallback enables pickle deserialization of .bin files.

SGLang contains an RCE vulnerability when the optional dumper subsystem is enabled, allowing for a sandbox escape when DUMPERSERVERPORT is set, enabling code execution on inference requests.

SGLang contains an unauthenticated RCE in /loadloraadapterfromtensors via bypass of SafeUnpickler’s incomplete denylist, allowing arbitrary command execution through crafted base64-encoded pickle payloads.

SGLangs multimodal generation runtime is vulnerable to an unauthenticated path traversal vulnerability, allowing an attacker to write arbitrary files anywhere the server process has write access, by including ../ sequences in the upload filename when sent to specific endpoints.

SGLang uses an expert-parallel backup subsystem that exposes a ZeroMQ PULL socket on a routable network interface that does not contain authentication or deserialization safeguards, allowing an attacker to provide a malicious pickle file that results in unauthenticated remote code execution when the feature is enabled and the service is reachable over the network.

SGLangs replayrequestdump.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.

A security flaw has been discovered in lmsys sglang 0.4.6. Affected by this vulnerability is the function main of the file /updateweightsfromtensor. The manipulation of the argument serializednamedtensors results in deserialization. The attack can be launched remotely. The exploit has been released to the public and may be exploited. The vendor was contacted early about this disclosure but did not respond in any way.

SGLang contains a model weight exfiltration vulnerability when no API keys are configured, as SGLang will expose two endpoints that allow a remote attacker to trigger distributed weight broadcasting using NCCL and then triggering data transfer, attackers can exfiltrate all model weights.

SGLang contains a credential leakage vulnerability in the /serverinfo endpoint, which will return API keys and SSL keyfile information when only the --admin-api-key is configured.

SGLang contains an SSRF and local file read in the multimodal generation endpoint /v1/chat/completions due to unsanitized imageurl, allowing access to internal metadata, secrets, and services.

A vulnerability was determined in sgl-project SGLang up to 0.5.11. Affected by this vulnerability is the function datahash of the component Cache Handler. This manipulation causes denial of service. The attack is restricted to local execution. A high degree of complexity is needed for the attack. The exploitation appears to be difficult. The exploit has been publicly disclosed and may be utilized. The pull request to fix this issue awaits acceptance.