When GPU Monitors Lie: A Cryptomining Case That Tests Cluster Trust
A reported Linux intrusion involving an exposed Ray job service, a kernel rootkit, and falsified NVIDIA telemetry shows how mining crews can aim to blind both host and hardware visibility.
High-value GPU systems are built on a simple assumption: if the scheduler looks normal and the telemetry looks normal, the cluster is probably normal. A reported operation tracked as REF3364 challenges that assumption by combining access to Linux GPU hosts with concealment layers designed to confuse both operators and monitoring tools.
Fast Facts
- REF3364 was reportedly tied to three Linux GPU hosts, though the total scope is not independently confirmed.
- The intrusion allegedly followed abuse of an internet-exposed Ray job agent, but the exact access path remains unverified.
- A kernel rootkit and falsified NVIDIA GPU telemetry were reportedly used to hide cryptocurrency mining activity.
- Tor and Ethereum were reportedly used as separate command channels, but their exact roles are not described.
- The material does not establish data theft, broader tenant impact, or responsibility by any vendor.
Why this matters
Ray’s security guidance warns that exposed Dashboard, Jobs, or Client services can permit reachable users to execute code on a cluster. That is a serious operational risk in any environment that treats remote job submission as a trusted path. If such services are left accessible beyond intended networks, they may be abused for unauthorized workloads, including cryptocurrency mining.
The concealment side is just as important. A Linux kernel rootkit can undermine host-local visibility by hiding selected processes, modules, or system artifacts. That does not make detection impossible, but it can weaken confidence in the very tools responders normally use first. MITRE ATT&CK treats rootkits as a recognized concealment technique, including in cases linked to cryptomining.
NVIDIA documents that nvidia-smi, through NVML, can surface multiple GPU state and process-related measurements. If those values are falsified, monitoring that relies on them may be misled. The key defensive lesson is narrow but important: a single host view is not enough when both the operating system and the GPU reporting layer may be untrustworthy.
The supplied material does not identify additional affected systems, data impact, or broader compromise beyond the three reported hosts. It also does not establish which telemetry paths were altered, how the manipulation was performed, or whether the concealment fully succeeded.
Defensive takeaway
For operators, the practical response is familiar but easy to neglect. Keep job-submission and dashboard services off the public internet where possible. Verify exposure, review access logs, and correlate GPU utilization with independent signals such as hypervisor telemetry, network connections, scheduler records, and kernel integrity checks. If root compromise is suspected, preserve evidence and assess from a trusted external environment rather than relying on local tools alone.
Conclusion
This case is less about one mining campaign than about a fragile trust model. When exposed orchestration services, host-level concealment, and manipulated GPU telemetry collide, defenders are asked to believe the machine that may already be lying. The broader lesson is blunt: cluster security depends on independent verification, not on a single dashboard telling a reassuring story.
TECHCROOK
Network firewall appliance: A dedicated firewall can help administrators restrict access to cluster job-submission and dashboard services to intended networks. Review its access rules alongside service configuration and logs; a firewall is one layer of network control, not a substitute for securing the services themselves.
WIKICROOK
- Rootkit: General background term for malware designed to hide itself or other malicious activity from standard system views.
- GPU Telemetry: General background term for measurements about GPU usage, memory, power, temperature, and related activity.
- Ray Job Agent: General background term for a Ray service that can handle job submission and cluster execution workflows.
- NVML: General background term for NVIDIA’s management library that underlies GPU monitoring tools such as nvidia-smi.
- Command Channel: General background term for a path used to send instructions to or receive responses from a compromised system.



