DIGGINGBEAGLE

Entity · DiggingBeagle record

SGLang

Open-source language-model serving framework with GPU KV-cache transfer and disaggregated inference support.

Record description

SGLang is an open-source inference-serving project whose architecture includes GPU-side request scheduling, cache management and optional disaggregated prefill/decode transfers. Its October 2026 technical account describes how asynchronous RDMA completion must be synchronized with cache-slot ownership. This is a project identity, not a claim that every SGLang deployment is affected by one vulnerability. The October 2026 coordinated disclosures distinguish a production rejection-path failure at one GLM-5.3 deployment from an independently reproduced RDMA cancellation-lifetime hazard. They share cache-isolation concerns without sharing an established victim count.

Claims & evidence

reported findingsupported

Cloudflare reports that its GLM-5.3 cross-customer incident came from an over-capacity prefill request entering decode, while the SGLang team separately reproduces an asynchronous RDMA cancellation/write-after-reassignment hazard. The two mechanisms must not be counted as the same confirmed production incident.

Declared connections

CaseSGLang accepts caller-supplied image hash as cache identity in native inference requestsCaseSGLang debug dumper executes caller-controlled Python inside its inference schedulerCaseCloudflare inference rejection path leaked another tenant’s cached contentCaseSGLang asynchronous RDMA transfers could overwrite a newly assigned user cacheEvidenceSGLang disaggregated-prefill KV-cache transfer safety researchEvidenceCloudflare GLM-5.3 cross-customer KV-cache exposure postmortem