Research

SANSAR Swaps libnvidia-ml.so.1 So nvidia-smi Reads 0%: Check Your NVIDIA Library Against Its Package and Close Your Ray Jobs API

SANSAR Swaps libnvidia-ml.so.1 So nvidia-smi Reads 0%: Check Your NVIDIA Library Against Its Package and Close Your Ray Jobs API

If you run Linux GPU hosts, do two things today. First, check that libnvidia-ml.so.1 on every GPU host matches the driver package that installed it: a crew Elastic Security Labs tracks as REF3364 replaced that file with a wrapper that makes nvidia-smi report 0% GPU utilization while the cards mine cryptocurrency. Second, if you run Ray, make sure its job submission API is not reachable from the internet without authentication. That is how this crew got in.

Elastic published the research on Thursday, November 12, 2026 ("0% GPU utilization: How SANSAR hides Linux cryptomining," by Salim Bitam). The Hacker News carried it on Monday, November 16. Everything below comes from Elastic's report unless we say otherwise. The population is small and stated plainly: Elastic describes one intrusion, deployed across three Linux GPU hosts, between Thursday, October 22 and Sunday, October 25, 2026.

What is confirmed

The way in was an open Ray job agent. Elastic: "The job agent remained reachable from the internet and accepted job submissions without authentication." The attacker probed, then submitted a job that ran nvidia-smi and sent the GPU inventory out, then submitted a Python reverse shell job, then added an SSH key to the local account's authorized_keys. In other words, they shopped for GPUs before they broke in.

Elastic does not name a CVE or a port. The Hacker News links the activity to the ShadowRay 2.0 campaign; that connection is the outlet's, not Elastic's. For background: Ray's unauthenticated Jobs API was assigned CVE-2023-48022 in 2023, and Anyscale, Ray's maintainer, disputes it as intended behavior for a cluster that is supposed to sit on a trusted network. We did not find that CVE in the CISA KEV catalog we sync. There is no patch to wait for. The fix is network exposure and authentication.

The NVIDIA library is replaced, not preloaded. The file swapped is libnvidia-ml.so.1, the NVIDIA Management Library (NVML) that nvidia-smi and most GPU monitoring tools read from. Elastic reports the Makefile finds the NVIDIA library directory with ldconfig and replaces that file with the wrapper, keeps the genuine versioned library in place so the wrapper can forward calls to it, and saves a backup at /usr/lib/.libnvidia-ml.so.1.backup if one does not exist. The wrapper's working files sit in /opt/vadi_/nvml-hook/. Elastic does not describe LD_PRELOAD.

What the wrapper lies about. It forwards each NVML call to the real library, then edits the answer: GPU utilization comes back as 0%, mining processes are filtered from the process list, temperature is replaced with a random value between 36 and 46 °C, and power with a random value between 100 and 200 W. Any tool that reads GPU state through NVML on that host gets the same doctored numbers, which is why a dashboard can look calm while the cards work flat out.

Persistence and hiding. After getting root, the attacker staged the toolkit under /opt/vadi_ and built a kernel module named acpi_thermal_mgr.ko through DKMS, plus an initramfs loader so it loads at boot. The module uses ftrace to hook getdents64, openat and kill, and to hide TCP and UDP socket entries; hidden items return "No such file or directory" or "No such process." By default it hides processes named kurt and lyrebird, and Elastic notes the module's self-hiding is disabled by default. A separate backdoor, netns, runs as a systemd service (Elastic gives no unit name).

The parts. sansar is a Rust orchestrator that talks to its C2 over Tor using Arti, checks TracerPid and exits if a debugger is attached, and starts and stops miners on command. kurt is XMRig 6.25.0 for CPUs. tilki is SRBMiner-MULTI for GPUs. lyrebird is the legitimate Tor Project pluggable transport (version 0.8.1), used to disguise the Tor traffic. netns reads a controller IPv4 address from an Ethereum smart contract, connects to it on TCP port 8080, and opens a reverse shell to whatever host and port the controller returns.

What is not known

Pools and wallets. Elastic publishes none. The C2 sends pool and wallet settings in its MinerStart messages, and Elastic notes the launch paths "provide no hardcoded fallback." So there is no pool to block and no wallet to trace from this report. We are not listing public mining pools; blocking those is a policy choice for your network, not an indicator of this crew.

How many victims. One intrusion, three hosts, is all that is published. Elastic does not describe wider targeting.

Who. Elastic found Turkish-language comments in the rootkit source. That is a clue about a developer, not an attribution, and Elastic does not attribute it.

Whether our edge saw Ray probes. We tried to run our edge check for Ray Jobs API paths and could not: the tool's credentials were unreachable from this session. So we make no statement either way about what our sensors saw.

What to do

1. Verify the NVIDIA library against its package. On Debian or Ubuntu, find the owner with dpkg -S libnvidia-ml.so.1, then run dpkg --verify on that package name; a 5 in the third column of the output means the file's checksum no longer matches the package. debsums -c on the package does the same job. On RHEL, Rocky or Fedora, run rpm -qf on the library path, then rpm -V on that package; again, a 5 in the flags means the digest changed. If the driver came from NVIDIA's .run installer rather than a package, there is no package database to check against: compute sha256sum of the library on a known-clean host with the same driver version and compare.

2. Look for the specific leftovers. The file /usr/lib/.libnvidia-ml.so.1.backup, the directory /opt/vadi_, a DKMS module named acpi_thermal_mgr (dkms status, lsmod), the same name inside your initramfs image (lsinitramfs on Debian-family systems, lsinitrd on RHEL-family), and a systemd service you do not recognize. One caution: the rootkit hides files, processes and sockets from a live system, so a clean result on a running host is weaker than a clean result from booting known-good media and checking the disk offline.

3. Measure power where NVML cannot lie. On an infected host, the utilization, temperature, power and process numbers that come through libnvidia-ml.so.1 are the attacker's numbers. Read power from outside the host instead: the server's BMC (for example ipmitool dcmi power reading), your rack PDU, or your cloud provider's metrics. A GPU host that reports 0% utilization through nvidia-smi while drawing near its full load at the wall is the tell. So is power that jumps around randomly between 100 and 200 W at 0% utilization.

4. Close the Ray door. Do not expose the Ray dashboard or Jobs API to the internet. Put it behind a VPN or an authenticating proxy, and restrict who on the internal network can submit jobs. Review your job history with ray job list for jobs you did not submit, especially ones that only run nvidia-smi, and anything that opens a socket back out. Check authorized_keys on every account Ray runs as.

5. Egress. GPU training hosts rarely have a business reason to speak Tor, to call public Ethereum RPC endpoints, or to open outbound connections to arbitrary hosts on TCP 8080. Default-deny egress on those hosts would have broken both the C2 and the backdoor here.

Elastic also published two YARA rules, Linux_Trojan_Sansar_11eea454 and Linux_Trojan_Netns_c94a4b17, in the report.

Indicators

From Elastic's report. These are Elastic's indicators, cited, not our observation.

Elastic also lists hashes for kurt (XMRig 6.25.0, SHA256 96fc528ca5e7d1c2b3add5e31b8797cb126f704976c8fbeaecdbf0aa4309ad46), tilki (SRBMiner-MULTI, SHA256 acc706afc9322a3eae1ef49049749e0a7c9f1a123895ac56dc8a289beb2fb5b7) and lyrebird (SHA256 cc1ad7f14777a098ccc91fd4e8f6143da14418c2b6fdf23b7f8f6b3d1b0b7f1d). We are keeping those three out of our attributed set. lyrebird is a legitimate Tor Project tool. The kurt hash is not unique to this crew: the same hash already sat in our index from a Cisco Talos report on Cisco Catalyst SD-WAN exploitation, ingested June 15, 2026. A stock miner build shows up wherever someone drops stock XMRig, so it says "miner," not "REF3364." It is still worth alerting on a GPU host that is not supposed to mine.

How we know

We read Elastic's report directly, and The Hacker News article for the ShadowRay framing. We looked up each indicator in our own index: Elastic's four attacker-built hashes and the onion address were already held, ingested on November 12, 2026 from Elastic's report by our vendor-research collector, without an actor or family label. That is redistribution of Elastic's work, not a detection of ours, and we claim no lead. We prepared attributed records for the four attacker-built hashes and the onion address (actor REF3364, family SANSAR or netns) under the source tag manual-batch-sansar-2026-11; that write could not complete from this session and will be retried.

Sources

  • Elastic Security Labs, "0% GPU utilization: How SANSAR hides Linux cryptomining," Salim Bitam, November 12, 2026: https://www.elastic.co/security-labs/threat-command/sansar-linux-cryptomining-malware
  • The Hacker News, "SANSAR Linux Cryptomining Malware Uses Modified NVIDIA Library to Hide GPU Activity," Ravie Lakshmanan, November 16, 2026: https://thehackernews.com/2026/11/sansar-linux-cryptomining-malware-uses.html
  • NVIDIA Management Library (NVML): https://developer.nvidia.com/management-library-nvml

Every indicator in this post is in the feed. Free.

1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.

Get your free key → analytics.dugganusa.com/stix/register