A financially motivated threat actor has been running a sophisticated cryptomining operation against Linux GPU servers, using a custom toolkit that routes commands through Tor, conceals mining activity by falsifying GPU sensor readings, and pulls its own controller's address from the Ethereum blockchain. Researchers at Elastic Security Labs, who published their findings November 16, identified the campaign as REF3364 and named the central malware component SANSAR.
The attack, observed between October 22 and 25, started at an exposed Ray job agent. Ray is an open-source distributed computing framework widely used in AI and machine learning environments, and its job API accepted submissions without authentication in the affected deployment. The Ray dashboard was restricted to loopback, but the job agent remained internet-reachable, giving the attacker a direct path to run arbitrary jobs on the target hosts.
The initial jobs ran `nvidia-smi` to collect GPU inventory from the compromised machines and send it outward, before switching to a Python-based reverse shell. Once that foothold was established, an SSH key was inserted into the account that owned the Ray session. The attacker then escalated to root and staged the full toolkit under `/opt/vadi_`.
The toolkit has six components. SANSAR, written in Rust and versioned at 0.1.3, acts as the orchestrator. It connects to a Tor onion service through Arti, Tor's Rust-based client implementation, and receives commands covering miner start and stop, configuration updates, interactive shells, and SYN traffic tasks. Authentication relies on an HMAC-SHA256 digest tied to an embedded key, and the binary checks for debuggers at startup by reading the TracerPid field and exiting if the value is non-zero. The miners SANSAR manages are XMRig 6.25.0 for CPU workloads and SRBMiner-MULTI 3.6.9 for GPU, with no hardcoded pool or wallet address in the binary itself.
The most technically elaborate piece of the toolkit is a modified NVIDIA Management Library wrapper. NVML is the library that tools like `nvidia-smi` call to query GPU utilization, temperature, and power draw. The malware swaps in a replacement `libnvidia-ml.so.1` that calls the legitimate library internally but intercepts results before they reach the caller. It reports GPU utilization as zero percent while substituting temperature readings with random values between 36 and 46 degrees Celsius and power draw figures between 100 and 200 watts. Any process tied to mining is filtered out of the response entirely, so standard monitoring tools display an idle card while it runs at full capacity.
Kernel-level concealment comes from `acpi_thermal_mgr.ko`, a rootkit loaded as a kernel module through DKMS with an initramfs hook that ensures it reloads after every reboot. It taps into the Linux ftrace subsystem to intercept syscalls including `getdents64`, `openat`, and `kill`, returning "No such file or directory" or "No such process" errors for files and processes it is configured to hide. The module parameters governing what gets hidden, including target process names and filename prefixes, are adjustable.
A separate backdoor called netns, registered as a systemd service, handles resilient command-and-control. Rather than storing its controller's IP address in the binary or on a fixed server, netns queries the Ethereum blockchain through public RPC endpoints. It sends calls to a hardcoded smart contract at address 0x43eb291f56688c63c93e8a93c4591fc997cf6349, reads an IPv4 address out of the contract state, and connects to that address on TCP port 8080. Once connected, it waits for a `SHELL <host> <port>` command that spawns a Bash session pointing to an arbitrary destination. Elastic noted the command loop has no authentication or integrity protection.
Turkish-language comments in the rootkit's source code offer a loose geographic signal, though Elastic stopped short of firm attribution.
The Ray framework has drawn similar attention before. In 2025, Oligo Security documented the ShadowRay 2.0 campaign, in which attackers exploited CVE-2023-48022 to submit malicious jobs to Ray clusters and spread miners across nodes using Ray's own scheduling features. At the time, researchers estimated more than 230,000 Ray servers were exposed to the internet.
Elastic's recommendations focus on restricting job submission access to GPU infrastructure, verifying that monitoring libraries have not been swapped by comparing `libnvidia-ml.so.1` against a known-good copy, auditing `authorized_keys` files and systemd units for unauthorized additions, and reviewing DKMS modules and initramfs images for unexpected entries. YARA rules and file hashes for all six components are included in the full report.
