- Elastic Security Labs says an attacker used an unauthenticated Ray job agent to deploy a mining toolkit across three Linux GPU hosts, then falsified GPU readings.
- The toolkit reported 0% GPU use, so teams that trust dashboards on a compromised host could see nothing wrong.
- Check which AI cluster interfaces face the internet, and confirm GPU and process data is also collected from outside the host.
The dial that tells you how hard your GPUs are working is only as honest as the machine it runs on. In an intrusion that Elastic Security Labs documented, that dial was replaced. It told operators their graphics cards were idle while a cryptominer used them.
What Elastic found
Elastic Security Labs is tracking the intrusion as REF3364. The toolkit combines custom malware with publicly available miners. Its core is SANSAR, a Rust program that manages CPU and GPU miners and takes its orders over Tor, the anonymity network.
The way in was ordinary. Ray is an open-source framework for spreading Python work, including AI training and inference, across many machines. Its job agent is the part that accepts work orders. On the first host, the internet could reach that agent, and it ran submitted jobs without asking who sent them. Elastic says the attacker obtained root access and deployed the toolkit across three Linux GPU hosts.
One detail matters for defenders. On the first host, Ray's dashboard listened only on the local loopback interface. The job agent did not. Locking down the part you look at did not lock down the part that runs commands.
Three days from probe to foothold
Elastic's records show a short timeline. Outside probing first appears in the Ray logs on October 22, 2026. Two days later, a job called nvidia-smi, NVIDIA's standard GPU status tool, and shipped the list of GPUs to an outside server.
On October 25, the actor submitted a Python job that opened a reverse shell. That is a connection from the victim machine back to a listener the attacker controls. The shell had the permissions of the local user who owned the Ray session. The actor then added an SSH key to that account, which gave persistent access.
How the readings were falsified
The deception works in two layers. First, the attacker replaced NVIDIA's management library, the code that monitoring tools call to read GPU data. Elastic says the replacement forwards each query to the genuine library, then changes selected results. Utilization comes back as 0%.
Heat and power get invented answers too. According to Elastic, the wrapper reports a temperature somewhere between 36 and 46 degrees Celsius and a power draw between 100 and 200 watts, chosen at random each time.
Second, a kernel module named acpi_thermal_mgr hides selected processes, files and network socket entries. The kernel is the core of the operating system, so ordinary tools that list processes ask it for the answer. Elastic says the module intercepts those requests and removes matching entries. Elastic also found a boot-time loader, so the module can load again after a restart.
The module's source contained Turkish comments. Elastic reports this as a detail of the code. It does not name or place the attacker.
A second way in, hidden on a public blockchain
The toolkit also includes a separate backdoor called netns. Elastic says it looks up its controller's address from an Ethereum smart contract through public Ethereum RPC services, then connects directly. The controller can then order it to open a shell to a chosen host. Elastic notes the command loop has no authentication or integrity check.
For its Tor connection, SANSAR builds in Arti, the Tor Project's Rust version of Tor, as a library. Elastic found a bridges file with seven WebTunnel entries and three obfs4 entries. Bridges are unlisted Tor entry points that help traffic get through where direct connections are blocked.
Elastic also compared the file fingerprint of the collected miner, called tilki, with the official SRBMiner-MULTI 3.6.9 Linux download. The two matched. So the miner is unmodified, publicly available software, and the abuse lies in how it was deployed.
The lesson: telemetry is a witness, not a guard
This is one intrusion on three hosts, not a measured trend. But it shows a design weakness worth fixing. Many teams spot a miner because the GPU runs hot or the power bill climbs. Here the attacker rewrote exactly those signals at the source.
The practical conclusion is that a number reported by a compromised host comes from the very system being questioned. It needs corroboration from somewhere that system cannot edit.
Questions to put to your team
Start with exposure. Which Ray job agents, and other cluster management interfaces, can be reached from the internet, and do they require authentication? Ask about the interface that runs jobs as well as the dashboard.
Then ask about independent evidence. Can the team compare GPU power and cooling figures from the facility or the cloud provider with what the host reports? Does any tool watch outbound connections, including Tor traffic, from GPU hosts? Do alerts fire on new kernel modules or changes to boot images?
If the answer to who can reach your GPU cluster is "anyone", the host's own dashboard should not be the only place you check.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Elastic Security Labs.




