SANSAR cryptominer fakes 0% GPU use on Linux AI hosts
Elastic Security Labs detailed SANSAR, a Rust cryptomining toolkit that hooks NVIDIA's NVML to report 0% GPU use and spreads via exposed Ray job agents on Linux AI infrastructure.
Elastic Security Labs documented a financially motivated cluster it tracks as REF3364, running a Linux cryptomining toolkit called SANSAR that tampers with GPU telemetry so monitoring tools report 0% utilization. In the observed intrusion, initial access was an exposed Ray job agent — the orchestration endpoint that fronts GPU clusters for AI and LLM workloads. The activity was seen in late October 2026 and reported by The Hacker News.
How it hides
SANSAR installs a modified NVML wrapper (libnvidia-ml.so.1, the NVIDIA Management Library) that hooks the calls tools like nvidia-smi use:
- GPU utilization is reported as 0%.
- Temperature is replaced with random values of 36–46 °C, power draw with 100–200 W.
- Process information is filtered.
The orchestrator (sansar) is written in Rust and receives commands over a Tor onion service via Arti, then runs whichever miner matches the host — kurt (XMRig 6.25.0) on CPU, tilki (SRBMiner-MULTI 3.6.9) on GPU. A second backdoor, netns, reads its controller's IPv4 address from an Ethereum smart contract, then opens a plain TCP reverse shell to that address on port 8080. Persistence is a ftrace-based kernel rootkit, acpi_thermal_mgr.ko, built via DKMS with an initramfs loader so it survives reboot; the toolkit stages under /opt/vadi_. SANSAR reads TracerPid and exits if it is non-zero.
Initial access
Per Elastic, the attacker compromised the first host through an exposed Ray job agent, submitted jobs running nvidia-smi to exfiltrate the GPU inventory, then ran a Python reverse shell, added an SSH key to authorized_keys, and escalated to root before staging the toolkit.
Detection
Indicators published verbatim by Elastic Security Labs:
# Elastic Security Labs — SANSAR / REF3364
Tor C2: efxisnqldelyr5ra25oqfrsut54b34g26e7zrsiuupsoekjof2gho2id.onion:9051
ETH C2: 0x43eb291f56688c63c93e8a93c4591fc997cf6349
SHA-256
sansar 00c9ac44291f83c72d54302e99d155a3e26d02225a3e516764d5833a99657c9e
kurt (XMRig) 96fc528ca5e7d1c2b3add5e31b8797cb126f704976c8fbeaecdbf0aa4309ad46
tilki (SRBMiner) acc706afc9322a3eae1ef49049749e0a7c9f1a123895ac56dc8a289beb2fb5b7
libnvidia-ml.so.1 675c24728f87f7aa0fa07527027c95c24c246049c65eaef89270ab04ea50a180
acpi_thermal_mgr.ko 6ba1aa275e0e2826e8b48b36473539ca56cbdc41f429e539a45f8532b084d7a4
netns df8ab6fd699e6ae90e63e9696bae8f01342708bafc2ed88af414337e4f237194
Elastic also ships YARA rules Linux_Trojan_Sansar_11eea454 and Linux_Trojan_Netns_c94a4b17 in the report's detection section.
Action checklist
- Don't expose Ray dashboards or job agents to the internet. Require authentication and network-isolate the cluster; an open Ray endpoint is remote command execution by design.
- Don't trust
nvidia-smialone on a suspect host. Cross-check GPU load against out-of-band power and thermal telemetry (IPMI, PDU, hypervisor) — the whole point of this toolkit is to make the in-band reading lie. - Hunt the IOCs above, plus
/opt/vadi_, an unexpectedacpi_thermal_mgrDKMS module, new initramfs loaders, and additions toauthorized_keys. - Block and alert on outbound Tor from GPU hosts.
Context
Elastic notes Turkish-language comments in the rootkit source but says it is unclear who is behind the operation — keep the attribution open. The pointed detail is the target: this is cryptomining built specifically for the AI-infrastructure boom, where idle GPU capacity is expensive, abundant, and increasingly fronted by orchestration services like Ray that operators forget are reachable.