Impact
SGLang’s encoder parallel disaggregation system deserializes untrusted data using pickle.loads() in the disaggregation module without authentication, which allows an attacker to supply malicious payloads. This results in unauthenticated remote code execution, granting full control over affected servers and compromising confidentiality, integrity, and availability. The vulnerability is classified as CWE-502: Deserialization of Untrusted Data.
Affected Systems
The affected product is the SGLang framework from the vendor SGLang. No specific version information is provided by the CNA; therefore, all releases that include the disaggregation module are potentially impacted. The vulnerable component is identified by the CPE cpe:2.3:a:lmsys:sglang:*:*:*:*:*:*:*:*.
Risk and Exploitability
The CVSS score of 9.8 indicates critical severity. The EPSS score is less than 1%, suggesting a low current probability of exploitation, and the vulnerability is not listed in the CISA KEV catalog. The exploit requires an attacker to send crafted data to the disaggregation module, and because it is unauthenticated and exposed over a network-accessible interface, the likely attack vector is remote network-based, inferred from the description.
OpenCVE Enrichment
Github GHSA