Actively exploited Security news
SANSAR: attackers mine crypto on Linux GPU servers while a fake NVIDIA library reports zero utilization
On 12 November 2026 Elastic Security Labs described an intrusion into Linux GPU servers used for AI workloads. The attacker got in through a Ray Jobs agent exposed without authentication, deployed cryptominers and replaced NVIDIA's monitoring library so that tools like nvidia-smi report 0% GPU utilization. Here is how it worked, the indicators, and what to check on your own servers.
Key points
- On 12 November 2026 Elastic Security Labs published an analysis of intrusion set REF3364 and SANSAR, a toolkit that mines cryptocurrency on Linux GPU servers.
- The attacker came in through a Ray Jobs agent exposed to the internet without authentication: a job running nvidia-smi, then a reverse-shell job, then an added SSH key and root.
- A wrapped libnvidia-ml.so.1 makes tools such as nvidia-smi report 0% GPU utilization, a temperature of 36–46 °C and 100–200 W of power draw, so mining does not stand out in monitoring.
- A kernel rootkit, acpi_thermal_mgr.ko, hides processes, files and sockets; command and control runs over Tor, and a second backdoor looks up its server address in an Ethereum smart contract.
- The report includes SHA-256 hashes, the C2 address, file paths and YARA rules.
What happened
On 12 November 2026 Elastic Security Labs researcher Salim Bitam published an analysis of an intrusion Elastic tracks as REF3364. The attacker mined cryptocurrency on three Linux GPU servers used for AI workloads and ran the operation with a custom Rust toolkit called SANSAR.
What stands out is how the mining was hidden. GPU miners usually give themselves away through utilization, temperature and power readings. So the attacker replaced NVIDIA's management library libnvidia-ml.so.1 (NVML), which nvidia-smi and most monitoring agents read from. The wrapper calls the genuine library and then rewrites the results: GPU utilization is always 0%, temperature is a random 36–46 °C and power draw a random 100–200 W. On dashboards the card looks idle.
Elastic does not attribute the activity to a known group and gives no victim count.
How the attack works
Initial access. The Ray dashboard listened only on loopback, but the job agent (Ray Jobs) was reachable from the internet and accepted submissions without authentication. Elastic says that between 22 and 25 October 2026 the attacker probed the service, submitted a job running nvidia-smi that sent the GPU inventory out, and then a Python reverse-shell job. They added their own SSH key to authorized_keys and, after getting root, deployed the toolkit.
Components:
sansar– a Rust client (v0.1.3) that manages the miners and talks to its C2 over Tor (the Arti library, directly or via bridges). It handles 16 message types, including interactive shells, miner control and SYN-flood tasks.kurt– XMRig 6.25.0 (CPU mining);tilki– SRBMiner-MULTI 3.6.9 (GPU mining, hash matches the official release). Pool and login settings come from the C2; nothing is hardcoded.lyrebird– a Tor pluggable-transport helper for obfs4 and WebTunnel bridges when direct Tor connections are blocked.acpi_thermal_mgr.ko– an ftrace-based kernel rootkit that hides processes, directories, files and sockets and can hide itself. It is built through DKMS and loaded from the initramfs at boot.netns– a separate backdoor running as a systemd service. It reads its C2 IPv4 address from an Ethereum smart contract via public RPC endpoints and connects to port 8080. The protocol has no authentication at all.
The client also resists debugging (a TracerPid check and a timing loop), stores its onion address encoded and suppresses shell history in the sessions it spawns.
Who is affected
The main risk is to Ray clusters and, more broadly, Linux servers with NVIDIA GPUs for AI that can be reached from the internet without authentication. Ray's documentation states that anyone who can reach the dashboard, Jobs or client ports can execute arbitrary code on the cluster. Token authentication has been available since Ray 2.52.0; per the docs it is on by default from 2.59 for local clusters and from 2.61 for all clusters.
The damage goes beyond stolen compute and a bigger cloud GPU bill. The attacker had root, a kernel rootkit and two backdoors, which means access to everything on those hosts, including models, data and credentials.
How to tell if you were hit
Indicators from Elastic's report (onion address defanged). Elastic lists no attacker IP addresses or wallets.
SHA-256:
00c9ac44291f83c72d54302e99d155a3e26d02225a3e516764d5833a99657c9e sansar
96fc528ca5e7d1c2b3add5e31b8797cb126f704976c8fbeaecdbf0aa4309ad46 kurt (XMRig 6.25.0)
acc706afc9322a3eae1ef49049749e0a7c9f1a123895ac56dc8a289beb2fb5b7 tilki (SRBMiner-MULTI)
cc1ad7f14777a098ccc91fd4e8f6143da14418c2b6fdf23b7f8f6b3d1b0b7f1d lyrebird 0.8.1
675c24728f87f7aa0fa07527027c95c24c246049c65eaef89270ab04ea50a180 libnvidia-ml.so.1 (wrapper)
6ba1aa275e0e2826e8b48b36473539ca56cbdc41f429e539a45f8532b084d7a4 acpi_thermal_mgr.ko
df8ab6fd699e6ae90e63e9696bae8f01342708bafc2ed88af414337e4f237194 netns
SANSAR C2 (Tor):
efxisnqldelyr5ra25oqfrsut54b34g26e7zrsiuupsoekjof2gho2id[.]onion:9051
Ethereum contract used by netns:
0x43eb291f56688c63c93e8a93c4591fc997cf6349 (selector 0x6ded96e7), connects to port 8080
Paths:
/opt/vadi_/
/opt/vadi_/tilki
/opt/vadi_/nvml-hook/
/usr/src/acpi_thermal_mgr-1.0/acpi_thermal_mgr.c
/usr/lib/.libnvidia-ml.so.1.backup
arti_data/state, arti_data/cache, bridges, .client_id (next to the executable)
YARA rules (Elastic):
Linux_Trojan_Sansar_11eea454
Linux_Trojan_Netns_c94a4b17
What else to check:
- Ray job history – jobs that run
nvidia-smi, a shell or an outbound connection, especially from unfamiliar addresses, - the NVML library – compare the hash of
libnvidia-ml.so.1with the driver package (dpkg -Vorrpm -V, depending on the distribution); a.libnvidia-ml.so.1.backupfile is a red flag, - kernel modules –
acpi_thermal_mgrin DKMS (dkms status) and its sources under/usr/src. Because the rootkit can hide itself, do not trust the running system; inspecting the disk from another system is more reliable, - unknown keys in
authorized_keysand systemd services you did not create, - outbound traffic to Tor, public Ethereum RPC endpoints and public-IP lookup services (ipify, ifconfig.me, icanhazip, checkip.amazonaws.com),
- power draw – measurements from the PDU, IPMI or hypervisor, which a library inside the OS cannot fake.
You can check file hashes in our hash calculator and pull indicators out of logs with the IOC extractor; both run entirely in your browser.
What to do
- Take Ray off the internet. The dashboard, Jobs and client ports belong on an isolated network or behind an authenticating proxy. Ray's documentation stresses that TLS is no substitute for network isolation.
- Turn on token authentication (Ray 2.52.0 and later), ideally by upgrading to a release where it is the default. It is a safety net, not a replacement for isolation.
- Verify the NVIDIA libraries on every GPU host against the driver package and monitor them with file integrity checks.
- Rebuild compromised hosts from a clean image. A kernel rootkit loaded from the initramfs means manual cleanup is not enough. Rotate SSH keys and every credential and token that lived on the host.
- Alert on mismatches between reported GPU utilization and actual power draw or job throughput.
Timeline
- 22–25 Oct 2026 – the attacker probes Ray Jobs, submits nvidia-smi and reverse-shell jobs, adds an SSH key and deploys SANSAR (per Elastic).
- 12 Nov 2026 – Elastic Security Labs publishes its analysis with indicators and YARA rules.
Verified against sources on 16 November 2026.
FAQ
What is Ray and why was it the way in?
Ray is an open-source framework for distributed computing, widely used to train and serve AI models. Ray's own documentation warns that anyone who can reach the dashboard, Jobs or client ports can run arbitrary code on the cluster. Built-in token authentication arrived in Ray 2.52.0; per the docs it is on by default from 2.59 for local clusters and from 2.61 for all clusters.
How can I tell if nvidia-smi is lying?
According to Elastic, the wrapper always returns 0% utilization plus random temperature and power values in a narrow range. Be suspicious if a GPU shows zero load while jobs run slowly or power use measured outside the OS keeps rising. The reliable check is to compare the hash of libnvidia-ml.so.1 with the file shipped in the driver package and to look for /usr/lib/.libnvidia-ml.so.1.backup.
Is deleting the miner files enough?
No. The attacker had root, installed a kernel module that loads at boot, a systemd backdoor service and their own SSH key. Rebuild compromised hosts from a clean image, rotate every key and credential that lived on them (cloud, model hubs, data stores), and only then put them back into service.
Who is behind it?
Elastic does not name an actor or give a victim count; it describes one incident involving three hosts. The goal is financially motivated mining, although the client can also launch SYN-flood traffic on command.
Sources
Spotted a mistake? Let us know and we will correct it openly.