SANSAR: verifying GPU telemetry on Linux servers
SANSAR changes GPU readings after Linux servers are compromised. An enterprise response must establish who can submit jobs and where its usage evidence originates.
Portekizce okuIn this article
By Pentesty
What the investigation establishes
Elastic Security Labs published its analysis on November 12, 2026; Cyber Security Brazil covered it on November 16. Activity observed between October 22 and 25 began with an internet-reachable Ray job agent. According to Elastic, the attacker obtained root and deployed the toolkit across three Linux GPU hosts.
SANSAR coordinates CPU and GPU miners over Tor. The toolkit includes a rootkit, boot persistence and a modified NVML library that reports GPU utilization as 0%. This defines the scope of the published observation: a documented intrusion whose relevance to other environments requires evidence from those environments.
Verify the entire path to execution
In the documented case, the dashboard listened only on loopback while the job agent remained exposed. An enterprise review should map the services reachable on every node, inbound rules and the identity permitted to execute code. A screenshot of a protected dashboard covers only part of that chain.
The Ray security documentation recommends controlled networks, trusted code and separate clusters where isolation is required. It also documents token authentication as an additional layer. Record the installed version and check the effective configuration: availability and defaults for this feature vary between versions.
Separate the evidence from the component under investigation
When a query passes through an altered library, collecting the same reading through another dashboard can reproduce the problem. Identify which agents depend on local NVML and who can replace it. Compare files and package references from a trusted environment; preserve the results outside the server under investigation.
Cross-check authorized jobs against records outside the host, such as archived submissions, access decisions and available infrastructure-layer metrics. Independence depends on where each record originates. Reserved GPU billing helps investigate costs, but alone it establishes neither instantaneous utilization nor unauthorized mining.
Turn the suspicion into a recovery plan
Preserve evidence and agree isolation boundaries with the workload owner before changing the server. Review SSH keys, services, kernel modules, boot images and access to secrets reachable by the compromised account. The report describes an initramfs loader; the investigation must follow the boot chain.
For hosts whose integrity has been lost, plan reconstruction from trusted images and review exposed credentials according to the established scope. Record the image provenance, access changes and subsequent validation. Our analysis of Linux host recovery explains why closure depends on this evidence.
Define acceptance criteria the operations team can repeat
In an isolated replica, run an authorized workload with an owner, a time window and an expected result. Compare the recorded job, allocated capacity and telemetry sources. Also perform a benign change to a monitored file and confirm that the integrity alert reaches the external system responsible for tracking it.
Acceptance should bring together a service inventory, an access test from an unauthorized origin, integrity evidence and legitimate execution after remediation. Assign an owner to review each result and decide when to repeat the assessment. Reviewing container permissions and host access complements this scope on shared platforms.
Frequently asked questions
Does 0% GPU utilization establish that a server is free of mining?
Interpret that reading alongside its origin and integrity. In the SANSAR case, the library queried by monitoring tools was modified to report utilization as 0%.
Does protecting the Ray dashboard resolve job exposure?
Verify all execution paths and the nodes involved. The investigation describes a dashboard listening on loopback and a job agent that remained reachable from the internet.
Is a driver update enough to recover a compromised host?
The observed toolkit also uses boot persistence and a rootkit. Recovery needs to address system integrity, access and credentials within the investigated scope.
What evidence allows the assessment to close?
A verifiable set of access controls, image and boot integrity, legitimate jobs and telemetry with known origins. The service owner should approve the results and their limitations.
Sources and references
- Cyber Security Brazil: editorial news reference
- Elastic Security Labs: SANSAR investigation, published November 12, 2026
- Ray: cluster security model, isolation and authentication
Original Pentesty analysis, published November 16, 2026. The proposed scenarios are editorial guidance, with scope and validation to be defined for each environment.