As reported by The Hacker News, Elastic Security Labs has detailed a financially motivated threat cluster designated REF3364, which deploys a sophisticated Linux cryptomining toolkit called SANSAR. What sets this campaign apart from the endless parade of coinminer variants is the deliberate engineering that went into evading detection across every layer of the stack — from boot persistence to falsified hardware telemetry.

Threat Alert: As reported by The Hacker News, Elastic Security Labs has detailed a financially motivated threat cluster designated REF3364, which deploys a sophisticated Linux cryptomining toolkit called SANSAR.

Why This Matters Beyond Cryptomining

Cryptomining is often dismissed as a nuisance-level threat — a resource thief that costs electricity but doesn't exfiltrate data. SANSAR challenges that complacency. The toolkit includes a kernel rootkit with initramfs persistence, a TOR-based C2 channel using pluggable transports, and critically, a backdoor ("netns") that retrieves its command-and-control address from an Ethereum smart contract. That last detail is especially significant: it represents a growing trend of threat actors abusing decentralized, immutable infrastructure for C2 resolution, making takedown operations effectively impossible through traditional legal or ISP-level coordination.

The initial access vector — an exposed Ray job agent — is also notable. Ray is a distributed computing framework widely used in machine learning and AI workloads. As organizations race to deploy AI infrastructure, misconfigured or internet-exposed orchestration frameworks are becoming a preferred entry point for threat actors targeting GPU-rich environments. The threat actor used Ray's job submission interface to run nvidia-smi for hardware reconnaissance before deploying a Python reverse shell for persistent access.

The Telemetry Evasion Problem

The modified NVML wrapper hooks calls used by monitoring tools like nvidia-smi to report GPU utilization as 0%, while substituting temperature and power readings with randomized but plausible values.

This is the most technically concerning aspect of the campaign. Most GPU-based cryptomining detection relies on monitoring tools that query NVML — the same library that nvidia-smi, Prometheus exporters, and DCGM-based dashboards use. By intercepting and falsifying those calls, SANSAR ensures that every standard monitoring pipeline reports normal operation while the GPU is actually mining at full capacity. Security teams that rely solely on GPU telemetry for mining detection are effectively blind to this threat.

Attack Chain Summary

StageTechnique
Initial AccessExposed Ray job agent — command execution via job submission
Reconnaissancenvidia-smi via Ray job to enumerate GPU inventory
PersistenceSSH key injection; kernel rootkit via DKMS + initramfs loader
EvasionModified NVML wrapper falsifying GPU utilization, temperature, and power readings
C2TOR with pluggable transports; Ethereum smart contract for C2 address resolution
ImpactXMRig (CPU) and SRBMiner-MULTI (GPU) cryptomining; SYN flood capability

Who Is at Risk

The Telemetry Evasion Problem
AI/ML teams running Ray or similar distributed frameworks with job agents exposed to the internet
Cloud GPU instances (AWS p3/g4, GCP A2, Azure NC-series) where monitoring relies on in-guest NVML queries
Bare-metal GPU clusters in research or HPC environments with shared multi-tenant access
Cryptocurrency-adjacent infrastructure already running high GPU utilization, where anomalous mining blends in naturally

Shield53 Recommendations

Immediate Actions

  • Audit Ray deployments: Ensure Ray dashboard and job agent ports (default 8265, 10001) are not exposed to the internet. Restrict to internal networks or VPN access only.
  • Verify NVML integrity: Check for unexpected libnvidia-ml.so files or LD_PRELOAD entries. Compare nvidia-smi output against independent metrics sources such as DCGM exporter or cloud provider's hypervisor-level monitoring.
  • Hunt for SANSAR artifacts: Look for files under /opt/vadi_*, kernel modules named acpi_thermal_mgr, and processes named sansar, kurt, tilki, or lyrebird. Check authorized_keys for unauthorized SSH keys on Ray service accounts.
  • Inspect initramfs: Verify initramfs images for unexpected DKMS modules using lsinitramfs or dracut --list-modules.

Strategic Defenses

  • Independent telemetry: Never rely solely on in-guest GPU monitoring. Cross-reference with hypervisor-level or hardware BMC metrics that cannot be hooked by guest-level libraries.
  • Network egress controls: TOR pluggable transports and Ethereum RPC calls to public nodes are network-detectable. Block or alert on unexpected outbound TOR connections and cryptocurrency RPC endpoints from GPU compute nodes.
  • Ray hardening: Follow the Ray security documentation — enable authentication, restrict job submission APIs, and segment Ray clusters from general network access.
  • Ethereum smart contract monitoring: While difficult to implement broadly, organizations should be aware that blockchain-based C2 resolution is an emerging pattern. Threat intel feeds are beginning to track known smart contract addresses used for this purpose.

The SANSAR toolkit represents a maturation of Linux cryptomining operations — moving from opportunistic script-kiddie deployments to engineered, multi-layered persistence and evasion. For organizations investing heavily in GPU infrastructure for AI workloads, this campaign is a reminder that those same assets are high-value targets for financially motivated adversaries who are investing in the tooling to exploit them undetected.