SecurityOnlineDo Son
SANSAR malware hijacks AI GPU servers to mine cryptocurrency
Elastic Security Labs identified an intrusion cluster, tracked as REF3364, that deployed SANSAR Linux malware across three GPU servers used for AI workloads. The attackers first exploited an unauthenticated Ray job agent, which remained publicly reachable, to run commands.
Probes show a sequence of activity from 22 to 25 October 2026, including an nvidia-smi query and GPU inventory exfiltration, followed by a Python reverse shell and the addition of an SSH public key to a local account before gaining root access to propagate the toolkit to three Linux GPU hosts. The operation culminated in a Rust-based orchestrator, a kernel rootkit and a GPU telemetry spoofing library, plus an Ethereum-backed backdoor, all designed to mask cryptomining activity.
The threat uses several explicit components: CPU mining with XMRig 6.25.0 under the name kurt, GPU mining with SRBMiner-MULTI 3.6.9 under the name tilki, and a modified NVIDIA Management Library wrapper (libnvidia-ml.so.1) to disguise GPU utilisation. A kernel module named acpi_thermal_mgr.ko, deployed via DKMS with an initramfs loader, hides processes and network artefacts and blocks terminations unless the sansar manager process is active.
The Rust client, sansar (v0.1.3), connects to a Tor controller on port 9051 using Arti, then authenticates and coordinates miners and interactive shells. An ancillary systemd service backdoor named netns communicates with a hardcoded Ethereum contract to obtain an encoded IPv4 address, enabling outbound reverse shells to operator-controlled hosts.
Defenders are advised to inspect for /opt/vadi_ staging folders, unauthorized DKMS sources under /usr/src, and suspicious libnvidia-ml.so.1 integrity and outbound Tor/Ethereum activity.