Tuesday, 02 January 2024 12:17 GMT

Hackers Falsify NVIDIA GPU Readings To Mask Cryptomining Arabian Post


Date:

(MENAFN- The Arabian Post) clearfix"> Hackers concealed cryptocurrency mining on compromised Linux servers by manipulating NVIDIA graphics processor readings and hiding malicious processes, according to an investigation published by Elastic Security Labs on Thursday.

The intrusion, designated REF3364, affected three Linux hosts equipped with graphics processing units (GPUs). Attackers combined a kernel rootkit, altered monitoring software and two separate communication mechanisms to maintain access while disguising unauthorised computing activity.

Elastic researchers identified an internet-accessible Ray job agent as the initial entry point. Ray is software used to distribute computing tasks, including artificial intelligence training and inference, across multiple machines. The compromised agent accepted job submissions without authentication, allowing commands to run remotely.

The findings highlight a particular vulnerability for organisations operating costly GPU infrastructure: ordinary monitoring utilities can display misleading information when attackers interfere with the libraries supplying their measurements. The investigation did not establish that the operation extended beyond the three identified hosts.

Evidence collected by Elastic dates the first external probes to 22 October. On 24 October, an attacker submitted a job executing NVIDIA's nvidia-smi utility and transmitted information about installed GPUs to an external server. A Python reverse shell submitted the following day provided interactive access.

The initial shell operated with the permissions of the account running the Ray session. The attacker subsequently installed an SSH authentication key for persistent access and obtained root privileges. Researchers did not establish precisely how the privilege escalation occurred.

Once privileged, the operator installed a collection of malicious components under a directory named /opt/vadi_. Central to the operation was SANSAR, a programme written in Rust that coordinated cryptocurrency miners and communicated with its controller through the Tor anonymity network.

See also Qilin and Settra exploit patched Palo Alto VPN flaw

The software supervised both processor-based and GPU-based mining. Its CPU component used XMRig, stored under the name kurt, while the GPU component used SRBMiner-MULTI, renamed tilki. Mining pool configurations and wallet information could be supplied remotely, rather than embedded permanently in the executables.

A modified version of NVIDIA's Management Library, known as NVML, supplied the principal deception. Instead of disabling legitimate monitoring functions, the replacement library passed requests to the genuine software and selectively changed the responses returned to administrators and monitoring applications.

GPU utilisation could consequently appear as zero even when mining activity was underway. The altered library also generated temperature readings between 36 and 46 degrees Celsius and power measurements between 100 and 200 watts, while filtering information about running GPU processes.

That approach undermines reliance on nvidia-smi and other applications using the same library. A seemingly idle graphics processor is therefore not sufficient evidence that a machine is free of unauthorised workloads, particularly when its monitoring components have been modified.

Another component, a Linux kernel module named acpithermalmgr, used tracing hooks to conceal selected processes, files and network socket entries. It could interfere with operating-system inspection and cause attempts to examine hidden objects to return errors suggesting they did not exist.

The attackers installed the module using Dynamic Kernel Module Support and configured an early-boot loader, enabling it to return after a restart. This persistence mechanism complicated detection through routine process listings and file-system checks.

SANSAR maintained its principal command channel through Tor, supporting remote miner configuration and interactive shell sessions. A separate backdoor, named netns, obtained a controller address from an Ethereum smart contract before establishing a direct network connection. Ethereum served as an address-discovery mechanism, not as the transport for remote shell commands.

See also AI features in 84% of surveyed ransomware incidents

Elastic identified evidence of GPU miner execution on two hosts, including a kernel log on one machine and GPU-related artefacts on another. The investigation did not disclose the identities of the affected organisations, the amount of cryptocurrency generated or the financial cost of the intrusion.

The researchers urged operators to restrict access to Ray job-submission interfaces and examine changes to NVIDIA libraries, kernel modules, startup components and SSH keys. Such checks are particularly relevant where monitoring output conflicts with other indications of computing activity.

The Ray dashboard on the initial host was bound to its local interface, but that restriction did not protect the separately exposed job agent. The distinction matters because administrators could mistakenly assume that limiting dashboard access also prevented outsiders from submitting executable workloads.

Elastic also published detection material covering the SANSAR orchestrator and netns backdoor, including malware signatures and technical indicators intended to help security teams identify the toolkit on Linux systems.

MENAFN15112026000152002308ID1111963303


The Arabian Post

Legal Disclaimer:
MENAFN provides the information “as is” without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the provider above.



More Story

Search