Your GPU is at 0 per cent and the fans are quiet, according to the monitoring tools on screen. In reality, a Rust-based malware toolkit is mining cryptocurrency on the card and rewriting the telemetry to hide it. Elastic Security Labs has documented SANSAR, a Linux cryptomining operation that deploys a modified NVIDIA Management Library wrapper to fake GPU readings, a kernel rootkit to hide processes, and a backdoor that pulls its command-and-control address from an Ethereum smart contract. Here is the full attack chain and how to defend Linux servers against it.
What Is SANSAR Malware
SANSAR is the name of the orchestrator at the centre of the toolkit, tracked by Elastic Security Labs as threat cluster REF3364. It is written in Rust, receives commands over the TOR anonymity network and manages both CPU and GPU miners on infected hosts. The activity was observed in late October 2026.
The full toolkit is built for stealth, not just profit. The components Elastic identified:
- sansar (the orchestrator): TOR-based command and control, launching and stopping the miner, configuration updates, spawning interactive shells, a TCP SYN flood task, and maintenance of the NVML hook
- kurt (XMRig version 6.25.0): the CPU miner
- tilki (SRBMiner-MULTI version 3.6.9): the GPU miner
- lyrebird: a legitimate TOR pluggable transport proxy, abused for C2 connectivity
- acpi_thermal_mgr.ko: a Linux kernel rootkit that hides processes, directories and sockets
- netns: a backdoor that retrieves its C2 address from an Ethereum smart contract before connecting directly
How the Attack Happens
Initial access came through an exposed Ray job agent. Ray is a popular framework for scaling AI and Python workloads, and its job agent can execute jobs on the cluster. That is exactly what the attacker did:
- GPU recon. The attacker submitted jobs running nvidia-smi to capture the GPU inventory and exfiltrated it to an external server.
- Reverse shell. A Python reverse shell job connected back to a remote listener.
- Persistence. An SSH key was added to the compromised account’s authorised_keys file.
- Root and staging. After obtaining root, the toolkit was staged under /opt/vadi_.
- Rootkit installation. The acpi_thermal_mgr kernel module was built through Dynamic Kernel Module Support (DKMS) and paired with an initramfs loader. Because the module and loader live in the boot image on disk, the rootkit reloads every time the host boots from that image.
SANSAR itself starts with an anti-debugging check. It reads the value of TracerPid, the field that shows the process ID of whatever is currently tracing or debugging it, and exits if the value is non-zero. After that it reaches out to an onion service for C2 communication over Arti, sending back host and GPU information and fetching miner status. Its main job is supervising the mining process and launching the right miner for the hardware.
How the NVIDIA Library Trick Works
The standout technique is the modified NVIDIA Management Library wrapper. NVML is the library that monitoring tools use to read GPU telemetry, and nvidia-smi is built on it. The attackers replaced the wrapper with one that hooks those NVML calls and returns false readings.
- GPU utilisation is always reported as 0 per cent, no matter how hard the miner pushes the card
- Temperature readings are swapped for random values between 36 and 46 degrees Celsius
- Power draw readings are swapped for random values between 100 and 200 watts
To anyone watching a dashboard, the GPU sits idle and cool. Physically, it is hashing at full load. Combined with the kernel rootkit hiding the miner process itself, the operation is designed to survive a casual look at standard monitoring.

Why This Matters Beyond Cryptomining
These techniques do not stay in the crypto-mining world. A modified library that falsifies telemetry, a rootkit that survives in the boot image, and a backdoor with C2 hidden in a blockchain are the same building blocks a more targeted attacker would use. The fact that this is a financially motivated operation running them on opportunistic victims means the tradecraft is getting cheaper and more automated every month.
The Ethereum-based C2 is worth a second look. By storing the C2 address in a smart contract, the operator can move the command server at will without touching the infected host, and defenders cannot simply blocklist a static IP or domain.
Who Should Worry About This Threat
The campaign is opportunistic rather than targeted, which widens the risk list considerably. Any organisation running GPU servers that face the internet through poorly guarded management interfaces is a candidate. That includes AI and machine learning teams running Ray clusters, rendering farms, research labs with shared CUDA boxes, and cryptocurrency-adjacent infrastructure already carrying high power bills that can hide a miner.
The use of DKMS to build the rootkit is a specific signal. DKMS exists so kernel modules can be rebuilt when the kernel updates, which makes it legitimate on GPU hosts that need driver modules. That same convenience is what lets the rootkit survive reboots and kernel upgrades quietly. If your patching process rebuilds modules automatically, it may also be rebuilding the attacker’s rootkit.
Elastic Security Labs published the technical details behind these findings, and security teams should read the report for indicators of compromise. For most Linux administrators, the takeaway is narrower: exposed compute dashboards are now entry points for automated attackers, and GPU telemetry can lie to you.
How to Defend Your Linux Servers
- Never expose Ray dashboards or job agents to the internet. That was the initial access vector. Put Ray behind authentication and bind it to internal interfaces.
- Audit authorised_keys regularly. An unexpected SSH key on an AI or compute account is a persistence signal.
- Watch for DKMS modules and initramfs changes. New kernel modules built through DKMS on production hosts deserve immediate investigation.
- Verify NVML and driver library integrity. If GPU telemetry reads 0 per cent under load, or temperatures look randomly generated, check the NVIDIA library files against the official packages.
- Monitor for TOR traffic and onion endpoints. Outbound connections to onion services from servers are almost never legitimate.
- Check /opt for unfamiliar staging directories. The toolkit staged itself under /opt/vadi_.
- Use file integrity monitoring on the boot image so rootkit loaders cannot land unnoticed.
- Correlate physical and reported metrics. Power draw at the wall versus what NVML reports will expose the fake wrapper quickly.
- Restrict outbound traffic from compute hosts. Miners need pool connections and the toolkit needs TOR. Egress filtering kills both.
- Inventory GPU hosts like production databases. These machines run expensive hardware and often hold credentials for model pipelines. Treat any unexpected process on them as an incident, not a nuisance.
What is SANSAR Linux malware?
SANSAR is a Rust-based orchestrator used by the REF3364 threat cluster for Linux cryptocurrency mining. It receives commands over TOR, manages CPU and GPU miners, and ships with a kernel rootkit, a modified NVIDIA library that fakes GPU readings and a backdoor that pulls its C2 address from an Ethereum smart contract.
How does SANSAR hide GPU mining activity?
It installs a modified NVIDIA Management Library (NVML) wrapper that hooks the calls monitoring tools use. GPU utilisation is reported as 0 per cent, while temperature and power readings are replaced with random values between 36 to 46 degrees and 100 to 200 watts, so dashboards show an idle card.
How did the attackers get into Linux servers?
Initial access came through an exposed Ray job agent. The attacker ran nvidia-smi jobs for recon, dropped a Python reverse shell, added an SSH key for persistence, and after gaining root staged the toolkit under /opt/vadi_ with a DKMS-built kernel rootkit.
How do I detect SANSAR on my server?
Look for unexplained DKMS modules and initramfs changes, unexpected SSH keys in authorised_keys, staging directories under /opt, outbound TOR or onion traffic, and GPU telemetry that reports zero utilisation under full load. Compare physical power draw against reported readings to spot the fake NVML wrapper.
