A memory-handling flaw in KVM, the hypervisor that ships inside the Linux kernel, allowed researchers to read roughly 127 GB of a cloud server's memory from inside a virtual machine they created themselves on DigitalOcean, The Hacker News reported. The bug is tracked as CVE-2026-99312, and fixed kernels have already shipped.
The practical stakes depend entirely on where the host is running. The flaw affects only hosts equipped with AMD processors and with nested virtualisation enabled — a configuration used to run VMs inside VMs. Intel hosts and AMD hosts running without nesting are out of scope, which means teams can prioritise a specific slice of the estate rather than remediating every KVM server in sight.
On shared infrastructure, host memory is not empty space. It is the layer where other tenants' guests live, so a guest that can reach beyond its own allocation can, in principle, touch data belonging to other customers. The source article does not itemise what was actually exposed; session tokens, encryption keys, database pages and application state are illustrative candidates for what host memory can contain, not a confirmed list from this incident.
Notably, no exploit code and no reports of active exploitation in the wild have surfaced for CVE-2026-99312. The research appears to be a controlled demonstration. That makes this a patching story, not an active-compromise story — but the size of the reachable buffer on multi-tenant hardware is reason enough not to defer it.
Why the patch is easy to miss
Because KVM is compiled into the kernel rather than installed as a standalone package, remediation is a kernel update followed by a reboot. Rolling out a patched image or installing new kernel packages does nothing for the running kernel until the host restarts — a step that regularly slips through maintenance slots on long-lived hypervisor nodes.
Distributions have begun shipping fixed kernels, but advisory identifiers and backport timing vary by vendor. Treat your distribution's own advisory — not a generic patch-level check — as the source of truth before you mark a host compliant.
Self-check: is nested virtualisation actually on?
Before you decide how wide the blast radius is, confirm which hosts are exposed:
cat /sys/module/kvm_amd/parameters/nested
A value of 1 means nested virtualisation is enabled on that AMD host and the host falls inside CVE-2026-99312's scope. A value of 0, or a missing kvm_amd module, means it does not. The setting is controlled by the nested=1 option passed to the AMD SVM kernel module, so it is worth verifying across the whole AMD fleet rather than spot-checking one or two nodes.
Recommended order of work
- Inventory hosts by CPU vendor and nested status, using the command above.
- Patch AMD hosts with nested virtualisation enabled first — kernel update plus reboot.
- Disable nested virtualisation (
nested=0) on AMD hosts where the capability is genuinely unused; that removes the exposure without waiting for a maintenance slot. - On any remaining unpatched multi-tenant AMD hosts, review access and hypervisor logs as a precaution. Nothing indicates compromise, but 127 GB of reachable host memory on shared hardware warrants the extra pass.
For Hong Kong teams running private clouds or hosted KVM clusters, the exposure is narrow but the remediation cost is low: a scoped inventory, a reboot, and a configuration flag.
據 The Hacker News 報道,Linux kernel 內建的 hypervisor KVM 存在一項記憶體處理漏洞,讓研究人員在 DigitalOcean 上自行建立的虛擬機內部,讀取了雲端伺服器約 127 GB 的記憶體。該漏洞編號為 CVE-2026-99312,修補 kernel 已經推出。
實際風險高低完全取決於主機的運行情況。此漏洞只影響配備 AMD 處理器而且啟用了 nested virtualisation(巢狀虛擬化)的主機 — 即在 VM 內部再運行 VM 的配置。Intel 主機以及未啟用巢狀虛擬化的 AMD 主機不在影響範圍之內,這意味著團隊可以針對特定範圍的機群優先處理,無需為所有 KVM server 逐一修補。
在共享基礎設施上,host memory 並非空白區域,而是其他租戶的 guest 所在的層級。因此,能夠超越自身配額範圍的 guest,理論上可以接觸到其他客戶的數據。原文並未逐一列明實際外洩的內容;session token、encryption key、資料庫頁面及應用程式狀態,只是 host memory 可能包含什麼的舉例,並非本次事件已確認的外洩清單。
值得注意的是,CVE-2026-99312 目前既未出現 exploit code,亦未有野外實際利用(in-the-wild exploitation)的報告。該研究看來是一次受控的示範。這意味著本事件屬於需要儘快修補的項目,而非正在發生的入侵事件 — 但在多租戶硬件上可觸及的 buffer 規模,已足以令人不宜拖延處理。
為何這項修補容易被忽略
由於 KVM 是編譯進 kernel 之內,而非以獨立 package 安裝,修復工作涉及 kernel 更新以及隨後的重新開機。推出已修補的映像或安裝新 kernel package,在主機重新啟動之前,對正在運行的 kernel 毫無作用 — 而這一步驟在長時間運作的 hypervisor 節點的維護時段中經常被遺漏。
各大發行版已開始提供修好的 kernel,但 advisory 編號及 backport 時間因廠商而異。在將主機標記為合規之前,應以你所用發行版自身的 advisory 為準 — 而非只作一般性的 patch level 檢查。
自查:巢狀虛擬化是否真的已啟用?
在決定影響範圍之前,先確認哪些主機受到影響:
cat /sys/module/kvm_amd/parameters/nested
數值為 1,表示該 AMD 主機已啟用巢狀虛擬化,該主機屬 CVE-2026-99312 的影響範圍。數值為 0,或 kvm_amd 模組不存在,則表示不受影響。此設定由傳予 AMD SVM kernel module 的 nested=1 選項控制,因此值得就整個 AMD 機群逐一驗證,而非只抽查一兩個節點。
建議的處理次序
- 使用上述指令,按 CPU 廠商及 nested 狀態建立主機清單。
- 優先修補已啟用巢狀虛擬化的 AMD 主機 — kernel 更新加上重新開機。
- 對確實未使用該功能的 AMD 主機停用巢狀虛擬化(
nested=0),這樣毋須等待維護時段即可消除風險。 - 對其餘尚未修補的多租戶 AMD 主機,謹慎起見檢視存取紀錄及 hypervisor 日誌。目前並無跡象顯示系統被入侵,但在共享硬件上有 127 GB 可觸及的 host memory,值得多做一次檢查。
對於在香港營運私有雲或代管 KVM 叢集的團隊而言,暴露範圍雖然有限,但修復成本很低:一份有範圍界定的主機清單、一次重新開機,以及一個設定旗標。
