A financially motivated threat cluster tracked as REF3364 has been linked to a Linux-focused cryptocurrency mining operation that takes an unusual step to stay hidden: rather than disguising its own programs or scheduling work for quiet hours, it tampers with the very library security teams query to detect heavy GPU load.
The campaign, reported by The Hacker News in November 2026, centres on a toolkit called SANSAR — a Rust-based orchestrator that accepts commands over the Tor anonymity network to manage both CPU and GPU miners on compromised hosts.
Tampering with the measurement layer
The core tradecraft is a modified copy of what appears to be NVIDIA's management library — an NVML-style interface that administrators and monitoring stacks query for GPU utilisation, power draw and temperature. By altering what that library reports, the operators effectively spoof the telemetry that underpins a large class of GPU-cryptojacking detection rules.
That marks an escalation beyond the usual anti-forensic tricks. Masking program names or deferring mining to off-peak windows merely makes activity harder to spot; rewriting the vendor's own readings breaks an entire detection methodology. Any alert logic built solely on "GPU utilisation exceeds X%" becomes unreliable against a host where the number itself is attacker-controlled.
The implication for defenders is direct: vendor telemetry can no longer be treated as the definitive record of what a machine is doing. Detection has to triangulate hardware-reported load against independent signals — power measured at the wall, PCIe throughput, thermal and fan behaviour, and process-level compute accounting that does not depend on the driver's self-reporting.
A resilient control plane in Rust and Tor
The design of SANSAR itself signals operator maturity. Writing the orchestrator in Rust yields a single, portable, statically linked executable that runs across heterogeneous Linux fleets without dependency headaches, while the language's memory-safety properties and compiled nature raise the cost of reverse engineering for analysts.
Command-and-control over Tor compounds that friction: there is no stable network footprint to place on a blocklist, no fixed IP range or domain name to redirect, and outbound traffic blends into whatever Tor activity an environment already permits. The result is a control plane that is language-agnostic, resilient to takedown of individual endpoints, and difficult to profile from network monitoring alone.
Detection checklist for Linux administrators
The mitigations below are procedural rather than dependent on patching — which is exactly why they are easy to overlook. This tradecraft does not hinge on a vulnerability in the driver stack so much as it targets trust in vendor-supplied tooling, and the controls that follow reflect our own analysis rather than guidance issued by any vendor or agency:
- File-integrity monitoring on driver and CUDA-adjacent libraries, specifically
libnvidia*paths, alerting on unexpected modification or replacement outside a package-manager transaction. - Package signature validation before installing or updating GPU driver components, and investigation of any driver artefacts that did not arrive through a trusted channel.
- Monitoring for unexpected Tor outbound traffic from hosts that have no legitimate reason to reach the anonymity network.
- Cross-checking GPU telemetry against at least one independent measurement source — wall power, PCIe throughput or thermals — rather than trusting a single reporting path.
Commodity cryptomining has steadily shifted toward stealth, and this campaign shows the frontier moving from hiding the miner to hiding the evidence of mining. For any IT team running Linux GPU infrastructure — whether for AI, HPC or scientific workloads — the lesson is general rather than specific: assume the measurement layer can be compromised, and verify it accordingly.
The campaign underscores a broader point about GPU-side threats on Linux: integrity verification, not patching, is now the operative control.
一個以牟利為動機、編號 REF3364 的威脅組織,被發現與一項針對 Linux 的加密貨幣挖礦行動有關。該行動採取了一個不尋常的手段以保持隱蔽:它並非偽裝自身的程序,或把工作排程至低峰時段,而是直接篡改保安團隊用來偵測 GPU 高負載的函式庫本身。
這項行動由 The Hacker News 於 2026 年 11 月報道,核心是一套名為 SANSAR 的工具——一個以 Rust 編寫的 orchestrator(協調器),透過 Tor 匿名網絡接收指令,在被入侵的主機上管理 CPU 及 GPU 挖礦程序。
篡改量度層
其核心手法(tradecraft),是修改一份看來屬於 NVIDIA 管理函式庫的副本——一個 NVML 風格的介面(interface),系統管理員及監控平台會查詢它以取得 GPU 使用率、功耗及溫度。通過改寫該函式庫的回報內容,攻擊者實際上偽造了支撐一大類 GPU 加密貨幣挖礦劫持(cryptojacking)偵測規則的 telemetry(遙測數據)。
這標誌著行動已升級,超越一般反取證伎倆。遮蔽 program(程序)名稱,或把挖礦延後至低峰時段,僅僅令活動較難被察覺;而重寫廠商自身的讀數,則是瓦解了整套偵測方法。任何單純建立在「GPU 使用率超過 X%」之上的 alert(警報)邏輯,面對一個數值本身由攻擊者控制的主機,都會變得不可靠。
對防禦者而言,含義直接而明確:廠商 telemetry 已不能再被視為機器運作的事實基準(ground truth)。偵測必須把硬件報告的負載與獨立信號交叉比對——包括牆上電錶量度的功耗、PCIe 吞吐量、散熱及風扇行為,以及不依賴驅動程序自我回報的 process(作業系統層面程序)運算核算。
以 Rust 和 Tor 建構的韌性控制平面
SANSAR 本身的設計已顯示操作者的成熟度。以 Rust 編寫 orchestrator,可產生單一、可攜、statically linked(靜態連結)的 executable(可執行檔),在各種異構 Linux 機群上運行而毋須處理依賴(dependency)問題;同時該語言的 memory safety(記憶體安全)特性及編譯性質,亦提高了分析人員逆向工程的成本。
透過 Tor 進行 command-and-control(命令與控制)更令此障礙加劇:沒有穩定的網絡足跡可供列入 blocklist(封鎖名單),沒有固定 IP 範圍或 domain name(網域名稱)可被重定向,而出站流量(outbound traffic)則混入環境原本已允許的 Tor 活動之中。結果是一個與語言無關、能抵受單一端點被取締、且單憑網絡監控難以描畫輪廓的控制平面。
Linux 系統管理員偵測清單
以下緩解措施屬程序性質,而非依賴修補(patching)——正因如此才容易被忽略。這種手法並非倚靠驅動程序堆疊的漏洞,而是針對對廠商提供工具的信任;下列控制措施反映我們自身的分析,並非任何廠商或機構發出的指引:
- 對驅動程序及 CUDA 相關函式庫進行 file-integrity monitoring(檔案完整性監控),特別是
libnvidia*路徑,一旦在 package manager(套件管理器)交易以外出現預期之外的修改或取代,即發出警報。 - 安裝或更新 GPU 驅動程序元件前進行 package signature validation(套件簽章驗證),並調查任何並非經可信渠道取得的驅動程序檔案(artefacts)。
- 監察主機上預期之外的 Tor 出站流量,尤其是那些毫無正當理由需要連接匿名網絡的主機。
- 將 GPU telemetry 與至少一個獨立量度來源交叉核對——牆上功耗、PCIe 吞吐量或散熱數據——而非僅信任單一回報路徑。
通用型加密貨幣挖礦已逐步轉向隱匿化,而這項行動顯示,前沿已從「隱藏礦工」推進到「隱藏挖礦證據」。對於任何運行 Linux GPU 基礎設施的 IT 團隊——無論是 AI、HPC 或科學工作負載——教訓是普遍而非特定的:假定量度層可被入侵,並據此加以驗證。
這項行動突顯了 Linux 上 GPU 端威脅的一個更廣泛要點:完整性驗證(integrity verification),而非修補(patching),現已是最關鍵的控制手段。
