Cryptomining malware has long had a visibility problem from the attacker's perspective: the moment a compromised host's GPU pegs at 100%, monitoring tools and alerting pipelines tend to notice. A financially motivated cluster tracked as REF3364 appears to have solved that problem for Linux targets by tampering with the very readings administrators rely on to spot abuse. According to reporting from The Hacker News, the operation centers on a Rust-based toolkit called SANSAR, which orchestrates both CPU and GPU miners and communicates with its operators over the TOR anonymity network.
What SANSAR Does
SANSAR is best understood as a command-and-control orchestrator rather than a single payload. Written in Rust — a language increasingly favored by malware authors for its cross-compilation ease and memory-safety characteristics — it receives instructions over TOR and uses them to deploy and manage mining components across a compromised Linux host. The dual CPU/GPU capability matters: it lets operators maximize yield on whatever hardware they land on, from modest virtual machines to GPU-equipped workstations and servers.
The distinguishing feature, however, is anti-forensics rather than mining efficiency. REF3364 tampers with a modified NVIDIA library so that the telemetry used to detect heavy GPU workloads reports normal-looking activity. In practice, that means the usual first signal of a cryptomining compromise — sustained, unexplained GPU utilization — can be quietly suppressed, leaving defenders to rely on other indicators.
Why This Matters for Defenders
GPU monitoring is one of the most widely used heuristics for catching illicit miners, particularly in environments with expensive accelerator hardware. An implant that manipulates the library layer beneath those readings undermines a detection strategy many teams treat as reliable. The lesson is not that GPU monitoring is useless, but that it should never be the only control in place.
Because the tampering occurs at the library level rather than in kernel space, integrity checking of shared objects and package verification become meaningful detection opportunities. Comparing installed NVIDIA libraries against vendor-supplied hashes, and alerting on unexpected modifications, gives defenders a way to catch this class of manipulation directly.
Detection Opportunities Beyond GPU Metrics
Since the miners are managed through TOR, network telemetry remains valuable. Unusual outbound connections to TOR entry nodes, or TOR-related traffic from servers that have no business reason to use anonymity networks, are strong indicators. Host-based signals include:
- Unexpected process lineages — mining binaries spawned by unfamiliar parent processes, or Rust-compiled executables appearing outside expected paths.
- Persistence artifacts — new systemd units, cron entries, or shell profile modifications introduced around the time of suspected compromise.
- Package and library drift — modified shared objects in library paths, particularly anything touching GPU drivers.
- Resource anomalies that don't match telemetry — rising power draw, fan speeds, or thermal readings inconsistent with reported utilization.
That last point deserves emphasis. Hardware-level signals — BMC telemetry, IPMI sensors, thermal and power metrics — sit below the layer SANSAR manipulates. Correlating reported GPU utilization against actual power consumption can expose the discrepancy the malware is trying to create.
The Broader Trend
REF3364 fits a well-established pattern: financially motivated actors treating cloud and on-premises Linux infrastructure as a mining substrate, with steadily improving operational security. The shift toward Rust-based, modular orchestrators with anonymized command channels raises the bar for simple signature-based detection, while library-level tampering shows attackers targeting the observability stack itself rather than just the operating system.
For defenders, the practical takeaway is layered verification. Trust telemetry, but cross-check it against independent sources: file integrity monitoring for driver and library components, hardware sensors for physical resource consumption, network monitoring for anonymizing infrastructure, and behavioral analytics for process and persistence anomalies. Any single control can be subverted; the overlap is where detection survives.
Organizations running GPU-heavy Linux workloads — rendering farms, ML training clusters, research compute — should prioritize reviewing whether their monitoring would catch a host whose reported utilization looks normal while its power draw says otherwise. That gap is precisely what SANSAR was built to exploit.