SANSAR Cryptomining Malware Fakes GPU Readings to Hide on Linux Servers
A financially motivated group tracked as REF3364 is running a Linux cryptomining operation that tampers with the readings defenders use to spot heavy GPU use. Elastic Security Labs documented the activity, which was observed in late October 2026, and centres on a Rust-based toolkit called SANSAR. It takes commands over the TOR anonymity network and manages both CPU and GPU miners.
According to Elastic, the attackers got in through an exposed Ray job agent, which let them run commands. They first used the NVIDIA System Management Interface (nvidia-smi) to collect the GPU inventory and send it to an external server. They then submitted a Python reverse shell job, and added an SSH key to the account's authorized_keys file for persistent access. After gaining root, they installed a kernel rootkit and set it to reload at boot through the initramfs.
The concealment is the notable part. A modified NVIDIA Management Library (NVML) wrapper filters process information and hooks calls used by tools such as nvidia-smi, so GPU utilisation shows as 0%. Temperature and power readings are replaced with random values between 36 and 46 degrees Celsius and 100 to 200 watts. The rootkit, acpi_thermal_mgr.ko, hides processes, files and socket entries, and returns "No such file or directory" or "No such process" for hidden items.
A separate backdoor named netns runs as a systemd service. It reads a controller address from an Ethereum smart contract, then connects over TCP port 8080 to open a reverse shell. SANSAR also exits if it detects a debugger. Elastic does not know who is behind the campaign, but found Turkish language comments in the rootkit's source code. It warns that cryptojacking diverts compute from legitimate workloads and raises operating costs.
Why it matters
Teams that rely on nvidia-smi, utilisation graphs or temperature and power alerts to spot cryptojacking can be misled, because this malware falsifies exactly those readings. Anyone building GPU capacity for AI workloads with an open job scheduler is exposed, and the initial access control (who can submit jobs) matters more than host-level monitoring that the attacker can tamper with. Persistence through the boot image also means a simple process kill or reboot will not clear an infected host.
What to do
- Check who can submit jobs to Ray job agents and remove any internet exposure.
- Review authorized_keys files for SSH keys you did not add.
- Look for an /opt/vadi_ directory, the acpi_thermal_mgr kernel module and an unexpected netns systemd service.
- Verify NVML libraries have not been modified and compare GPU readings against power or billing data outside the host.