Attackers put Linux GPU hosts to work mining cryptocurrency, then altered the software that would normally reveal the workload. Elastic Security Labs documented three compromised GPU hosts in its November 12, 2026 report, describing a toolkit that combines the SANSAR mining orchestrator with falsified GPU telemetry, a kernel rootkit and a separate backdoor.
Elastic tracks the intrusion activity as REF3364. The entry point was an internet-accessible Ray job agent accepting unauthenticated submissions—a configuration relevant to organizations using Ray to distribute Python workloads, including AI training and inference. On the first compromised host, restricting the head-node dashboard to the local loopback interface had left the job agent exposed.
From a Ray job to persistent access
Ray’s Jobs API provides a way to submit commands for applications running on a cluster. The attacker used that capability first to inspect the hardware, then to obtain a shell.
Ray logs recorded external probing on October 22, 2026. Two days later, a submitted job invoked the NVIDIA System Management Interface, nvidia-smi, and transmitted the GPU inventory to an external server. On October 25, a Python reverse-shell job connected to an outside listener under the local account that owned the Ray session. Adding an SSH key to that account’s authorized_keys file gave the attacker persistent access.
With root privileges, the attacker installed the toolkit in /opt/vadi_. Installation included building the acpi_thermal_mgr kernel module through Dynamic Kernel Module Support (DKMS), replacing the NVIDIA Management Library (NVML), and registering the netns backdoor with systemd.
The rootkit also gained a route back into memory after a reboot. Its module and a loader were incorporated into the disk’s boot image: the loader runs through initramfs, the temporary filesystem held in RAM before Linux switches to its normal root filesystem. Booting with that image can therefore reload the rootkit.
GPU mining followed later on two hosts. On one, a kernel log recorded /opt/vadi_/tilki starting; on the other, newly created caches and GPU contexts supported execution.
Making a busy GPU appear idle
SANSAR is the controller-facing component, rather than the miner itself. Written in Rust and identified by its filename and sansar-client v0.1.3 version string, it selects a CPU or GPU miner from instructions returned by its command-and-control server.
For CPU mining, it launches kurt, identified as XMRig 6.25.0, which takes external pool configuration. The GPU executable, tilki, is SRBMiner-MULTI 3.6.9. Elastic verified that its SHA-256 hash matched the executable distributed in that miner’s official Linux release.
Concealment operates at two levels. The replacement library, nvml-hook/libnvidia-ml.so.1, suppresses process information and changes the telemetry delivered to monitoring software. The Hacker News, reporting Elastic’s findings, described NVML hooks that force tools such as nvidia-smi to show 0% GPU utilization. They also substitute random temperature readings between 36 and 46 degrees Celsius (96.8–114.8 degrees Fahrenheit) and power readings between 100 and 200 watts.
Below that library, acpi_thermal_mgr.ko uses Linux Function Tracer, or ftrace, to conceal selected processes, files, directories and socket entries. It also has a configurable watchdog. The Hacker News described interception of calls for directory listings, file access and process termination signals: matching entries are removed, or requests are rejected with No such file or directory or No such process responses.
The operation thus interferes with both workload visibility and the readings administrators might use to investigate it.
Two separate routes to remote control
Before establishing its Tor connection, SANSAR checks for analysis. It reads /proc/self/status and terminates when TracerPid is nonzero, meaning another process is tracing or debugging it. A second check times a 1,000-iteration loop; a result of 501 milliseconds or longer also causes it to exit.
Elastic recovered the controller hostname from 62 encoded bytes. Decoding rotates each byte right by (i % 7) + 1 and applies XOR with a repeating 32-byte mask:
plaintext[i] = ROR8(ciphertext[i], (i % 7) + 1) XOR mask[i % 32]
The decoded address and hardcoded port are:
efxisnqldelyr5ra25oqfrsut54b34g26e7zrsiuupsoekjof2gho2id.onion:9051
SANSAR embeds Arti, the Tor Project’s Rust implementation. It locates arti_data/state, arti_data/cache, bridges, .client_id and lyrebird relative to its executable directory.
There are two connection options. Direct mode uses Arti alone, with an initial bootstrap deadline of 60 seconds. Bridge mode allows 900 seconds and uses the legitimate Lyrebird transport helper, identified by build metadata as version 0.8.1. The collected bridge configuration held seven WebTunnel entries and three obfs4 entries. These unpublished relay addresses offer alternative entry points when direct Tor access is blocked; both modes ultimately reach the same onion controller.
Through that channel, SANSAR exchanges commands, host inventory and miner status. Its GPU-name collection command is:
nvidia-smi --query-gpu=name --format=csv,noheader
Client identification includes the stored .client_id and compiled version 0.1.3. Beyond starting and stopping miners, the command handlers support configuration changes, connection management, interactive shells, NVML-hook maintenance and an IPv4 TCP SYN traffic task directed at a specified address and port.
The systemd-managed netns backdoor provides another control route. It queries the embedded Ethereum contract 0x43eb291f56688c63c93e8a93c4591fc997cf6349 for a controller IPv4 address, then contacts that address directly over plain TCP on port 8080. The controller’s response supplies a destination host and port for a reverse shell.
The operator remains unidentified, although Elastic found Turkish-language comments in the rootkit source code.
Observables and checks for GPU administrators
Elastic published the following SHA-256 hashes in its November 12 research on the October 2026 intrusion. They identify collected toolkit components, including the publicly available miners and legitimate Tor transport helper used by the attacker:
sansar:00c9ac44291f83c72d54302e99d155a3e26d02225a3e516764d5833a99657c9ekurt:96fc528ca5e7d1c2b3add5e31b8797cb126f704976c8fbeaecdbf0aa4309ad46tilki:acc706afc9322a3eae1ef49049749e0a7c9f1a123895ac56dc8a289beb2fb5b7lyrebird:cc1ad7f14777a098ccc91fd4e8f6143da14418c2b6fdf23b7f8f6b3d1b0b7f1dlibnvidia-ml.so.1:675c24728f87f7aa0fa07527027c95c24c246049c65eaef89270ab04ea50a180acpi_thermal_mgr.ko:6ba1aa275e0e2826e8b48b36473539ca56cbdc41f429e539a45f8532b084d7a4netns:df8ab6fd699e6ae90e63e9696bae8f01342708bafc2ed88af414337e4f237194
Cryptojacking can consume capacity intended for legitimate workloads and raise operating costs. For organizations adding GPUs for AI and large language model workloads, Elastic recommends checking who can submit jobs and whether GPU monitoring libraries have been modified.




