You still treat cryptojacking as the noisy, low-skill crime that trips a GPU graph and a help-desk ticket. That assumption is now part of the attack. Huntress says unpatched GeoVision camera management software handed Windows footholds to cryptojackers at 13 organizations, most of them between November 8 and 10. On Linux, a financially motivated cluster tracked as REF3364 is running SANSAR, a Rust orchestrator that takes orders over Tor and tampers with the NVIDIA library readings your cybersecurity monitors use to spot a hot GPU. The miner is quiet because the telemetry is hostile. If your cyber security program still waits for a utilization spike, you are waiting for a signal the operator already forged.

Camera Software Is Already an Admin Console

GeoVision is a Taiwanese surveillance and access-control vendor with customers in 110 countries. That footprint is the point. Camera platforms sit on Windows boxes in wiring closets, plant floors, and “temporary” VMs that survived three refresh cycles. Nobody puts them in the patch meeting. Nobody assigns them an owner who can take an outage. Huntress watched the same pattern repeat: every intrusion in this wave started with a service that belongs to GeoVision’s software, then the box started mining.

Call the software what it is. A camera management suite that runs as a Windows service is a privileged agent with a network listener, a local account model, and a vendor update cadence that lags everything you actually care about. Unpatched code in that class is a side door with a logo on it. The victims spanned industries, which is another way of saying the targeting was opportunistic. If the console is on the internet, it is in scope for someone else’s weekend.

GeoVision surveillance camera hardware of the kind managed by the unpatched Windows software used in the cryptojacking wave
Huntress says GeoVision camera management software gave cryptojackers a Windows foothold at 13 organizations in a three-day window.

This is a bad look for every vendor that ships a privileged Windows service next to a camera and then treats CVE intake as a firmware afterthought. It is a worse look for you if that service is still bound to a public IP. A firewall rule that permits “the NVR” from anywhere is an enrollment form for whoever is scanning this week. brute-force noise against those consoles is constant even when a given wave starts from a buggy service rather than a guessed password. You already know how this ends: one forgotten listener, then a miner, then a very late ticket.

Shodan Is Doing the Inventory You Skipped

Help Net Security notes that a Shodan search of the GeoVision surface is part of the story, which should not surprise you. Attackers do not need your CMDB. They need a product fingerprint, a default port, and a patch gap. Your physical-security stack is an IT estate with worse hygiene and a political shield: “we cannot reboot the cameras.” That shield is why cryptojackers keep showing up here. They are not after your video. They are after a Windows host that is allowed to be ugly.

Treat every VMS, NVR, and badge console as a domain-adjacent workstation until you prove otherwise. If it can run a service as SYSTEM, it can run a miner as SYSTEM. If it can talk outbound, it can pull a payload. security hardening for this class of box is overdue in shops that spent five years hardening laptops and left the lobby camera server on a vendor default.

SANSAR Turned GPU Telemetry Into Attacker Cover

The Linux side of the same business model is more interesting, and more insulting to how you detect. REF3364’s toolkit, SANSAR, orchestrates CPU and GPU miners and takes commands over Tor. Central to the job is a modified NVIDIA library that tampers with the readings used to detect high GPU workloads. Your dashboard can show idle while the card hashes. threat detection that keys off nvidia-smi, vendor telemetry, or “GPU percent” as a first-class alert is now attacker-influenced evidence.

Read that again if you run inference boxes, render farms, or a single “data science” workstation that nobody wants to patch. Those hosts already run hot. You trained the SOC to ignore them. SANSAR is built for that indifference. A Rust orchestrator on Tor is not a script kiddie dropping xmrig in /tmp. It is a control plane for a fleet, with enough tradecraft to lie to the one metric operations still trusts.

NVIDIA GPU hardware associated with hidden cryptomining activity that tampers with library-level utilization readings
SANSAR tampers with NVIDIA library readings so a GPU miner can run without lighting the utilization graphs most teams still watch.

Tor as the command channel is a gift to anyone who still writes “block mining pools” and calls it threat-protection. Pool traffic is optional when the orchestrator already has a stealth path out. Your egress policy has to describe processes and libraries, not just destination categories. defense in depth here means you accept that the GPU counter can lie, the pool list can change, and the C2 can look like a privacy client a developer installed “for testing.”

Stop scoring this as a 2018 nuisance. Power bills still move. So do cloud GPU invoices. A hidden miner on a rented A100 is a finance event that incident response will inherit after FinOps forwards a screenshot. The technical tell moved down-stack, into libraries and unsigned modules sitting next to the real driver. If you only watch user-mode miners named like miners, you are collecting trophies while the orchestrator keeps the fleet.

Your Cybersecurity Stack Still Scores the Wrong Metric

Connect the two stories and the operational failure is obvious. On Windows, the initial access was “camera stuff,” so it lived outside the vulnerability SLA you apply to Exchange and vCenter. On Linux, the dwell signal was “GPU busy,” so the operator forged that signal. In both cases the crime is boring and cash-positive. Your program is still tuned for loud malware and branded ransomware notes. Cryptojacking pays rent in the gaps.

You do not need a new platform to close those gaps. You need a worklist that assumes management software is software, and that host telemetry is evidence an attacker can edit. Do the immediate work this week, then keep the boring loop on the calendar.

  1. Inventory every camera, NVR, VMS, and access-control console as a server. Record the vendor, the Windows or Linux build, the listening services, the account that runs them, and whether the host can reach the internet. If you cannot name an owner, the owner is you. Pull the box off public addressing the same day. Management interfaces belong on a jump network, behind the firewall, with MFA on the admin path. Disable or uninstall GeoVision services you do not use. Uninstall beats “we’ll patch later” when the vendor’s clock and yours already disagree.
  2. Stop treating GPU percent as gospel. Alert on unexplained process trees, new libraries next to the NVIDIA stack, unsigned modules loaded into driver-adjacent paths, persistence in systemd/cron, and sustained power or PCIe throughput that disagrees with the reported utilization. Compare nvidia-smi (or the vendor tool) against an out-of-band source: BMC power, data-center PDU, cloud billing, even a thermal sensor. A forged counter and an honest wattmeter will not match. That mismatch is the detection.
  3. Contain miners like footholds, not like junkware. Kill the process, then isolate the host. Capture the GeoVision service version, the dropped binaries, the Tor client or library, and the scheduled tasks before you reimage. Rotate local and service credentials. Check adjacent camera servers for the same listener. If the box was domain-joined, assume the miner had time to look around. Your incident response write-up should list initial access (vendor service, exposed console, stolen admin) rather than “crypto miner removed.”
  4. Put physical-security software on the same patch and hardening standard as your domain controllers’ neighbors. That means vulnerability scans, EDR enrollment, restricted outbound, and a change window that is allowed to reboot a camera server. If the business refuses the window, document the risk as accepted standing access. Quarterly, re-check Shodan and Censys for your own product banners. If a search engine can find the console, so can REF3364’s cousins.

Keep the ongoing loop small enough that it actually happens. Watch for new listeners on camera subnets. Diff NVIDIA-related libraries after driver updates. Review cloud GPU spend against job schedulers. Hunt Tor clients on boxes that have no business running them. brute-force logs on remaining admin portals should page someone; a quiet success against a GeoVision login is worse than a noisy failure. None of this requires a branded appliance. It requires you to stop granting camera vendors and GPU graphs a trust they have not earned.

Frequently Asked Questions

If GPU utilization looks normal, can we ignore miner hunting on Linux?
No. SANSAR’s design assumes you will do exactly that. Corroborate utilization with power draw, billing, PCIe throughput, and file integrity around the NVIDIA stack. A quiet graph is consistent with a clean host and with a tampered library. You need a second instrument.
Should camera systems sit on the corporate LAN if we “only use them for video”?
Only if you manage them like Windows or Linux servers: patched, inventoried, EDR-covered, and firewalled away from the internet. The GeoVision wave started in the vendor’s own service. Video function does not reduce privilege. If the console can run code, it belongs in the same patch SLA as anything else that can.
Is cryptojacking worth incident response, or do we just kill the process?
Treat it as a foothold with a monetization module attached. Kill and reimage, rotate credentials, and look for the same vendor service or library tampering on sibling hosts. The cashout method is mining. The access method is what you actually have to close.

Sources

Take Control of Your Server Security

Don't let brute-force attacks slow you down. Try IPBan Pro risk-free for 30 days.

Secure. Automated. Lightweight.

Stay up to date with the latest news, releases and more.

Take Control of Your Server Security

Don't let brute-force attacks slow you down. Try IPBan Pro risk-free for 30 days.

Secure. Automated. Lightweight.