What happened
NVD (VulDB) published 2026-09-24 a cluster of issues in kvcache-ai mooncake (distributed KV cache engine for LLM inference) up to 0.3.12/0.3.13.post1/0.3.14-rc1 with publicly disclosed exploit details: authorization bypass in the UnmountSegment RPC path handler via client_id/segment_id manipulation (CVE-2026-96762, CVSS 7.3 HIGH with PoC), improper access controls in ScopedSegmentAccess::MountSegment (CVE-2026-96763, CVSS 5.4), and resource-consumption / allocation issues in the MasterService::GetReplicaListByRegex regex handler (CVE-2026-96764, CVSS 4.9). Vendor was contacted but did not respond.
Why it matters
Mooncake is the KV-cache transfer/storage layer used to split prefill/decode across GPU clusters for LLM serving; remote authorization bypass in its RPC handlers lets an unauthenticated/remote caller unmount or alter other tenants' cached segments - a cross-tenant data-integrity and availability flaw specific to multi-tenant AI inference infrastructure.
Attack vector
Remote network requests to mooncake RPC path handlers (UnmountSegment/MountSegment) manipulating client_id/segment_id to bypass authorization, plus regex-based resource exhaustion.
Affected systems
kvcache-ai mooncake <= 0.3.12 / 0.3.13.post1 / 0.3.14-rc1
Mitigation
No vendor response/public fix; restrict RPC endpoint network exposure; monitor for the published PoC patterns per VulDB 409016.