SANSAR Linux malware fakes GPU use to hide mining

Hackers hid cryptocurrency mining on Linux servers running powerful graphics cards by falsifying the very readings administrators use to spot heavy workloads, according to research published by Elastic Security Labs on 12 November 2026. The toolkit, tracked as REF3364 and named SANSAR, combined a Rust-based controller, a kernel rootkit and a tampered NVIDIA monitoring library that reported GPU use as zero while mining carried on in the background.
The intrusion reached three Linux GPU hosts through an exposed Ray job agent. Ray is an open-source framework for distributing Python workloads, including AI model training and inference, across multiple machines. Its Jobs API lets users submit commands to run on a cluster. On the compromised hosts, the Ray dashboard on the head node listened only on the local loopback interface, but the job agent remained reachable from the internet and accepted job submissions without any authentication.
Recorded activity began with external probes on 22 October 2026. On 24 October, a submitted job ran NVIDIA's nvidia-smi utility and sent the GPU inventory to an external server. The next day, the attacker submitted a Python reverse-shell job, then added an SSH key to the account that owned the Ray session to keep access. After obtaining root — by a method Elastic's account does not explain — the operator staged the toolkit under /opt/vadi_/, built its kernel module and added an early-boot loader so the infection would survive a restart.
SANSAR itself is a Rust orchestrator, version 0.1.3, that receives commands over Tor using the Tor Project's Arti library, with Lyrebird bridges as an alternative route. It supervises publicly available mining software rather than mining directly: its CPU branch runs XMRig 6.25.0, renamed "kurt", while its GPU branch launches SRBMiner-MULTI 3.6.9, renamed "tilki". The controller supplies pool addresses, wallet or login details and other settings, and it can also open interactive Bash shells, change miner settings and generate TCP SYN traffic.
The deception at the heart of the campaign targets the NVIDIA Management Library, or NVML, the interface used by monitoring tools such as nvidia-smi. The attacker's wrapper forwards requests to the genuine NVIDIA library, then changes selected results before returning them. GPU utilisation is reported as 0%, while temperature readings are replaced with random values between 36 and 46 degrees Celsius and power figures with random values between 100 and 200 watts. SANSAR can even rebuild and reinstall the wrapper from recovered source files, keeping the false readings in place.
A separate kernel rootkit, acpi_thermal_mgr.ko, hides selected processes, files and socket entries by intercepting system calls, including directory listings. Its source code contains Turkish-language comments. A second backdoor, named netns, retrieves its controller address from an Ethereum smart contract before connecting directly over TCP, giving the attacker yet another route back into the machine.
Elastic's warning to defenders is blunt: a zero reading from an altered library cannot establish that a server is clean. Investigations should check who can submit Ray jobs, verify whether GPU monitoring libraries have been changed, and account for persistent access and kernel-level hiding rather than treating the visible miner as the whole infection. With AI training servers now among the most valuable targets on the internet, an exposed job interface is an open door — and SANSAR shows how thoroughly attackers can cover their tracks once they walk through it.