Case · DiggingBeagle record

An exposed Ray job agent became a route into GPU hosts whose reported utilization was later falsified

Elastic analyzed October activity against an unauthenticated Ray job-agent endpoint. The attacker inventoried NVIDIA hardware, introduced command execution and ran cryptomining components while deploying kernel and library-level concealment.

Public disclosure: Nov 12, 2026

Scope

A network-reachable Ray job-agent API accepts unauthenticated job work or commands on a GPU host.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    A network-reachable Ray job-agent API accepts unauthenticated job work or commands on a GPU host.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    The campaign used a replacement libnvidia-ml.so.1 to falsify process enumeration, load and temperatures, alongside a rootkit and Tor-capable implant. A separate on-chain Ethereum read supplied a changing controller address; Ethereum here served command-and-control discovery rather than proving wallet theft.

    Boundary: mechanism

  3. 03

    Observed consequence

    The compromise crosses from job admission into host execution and then from host compromise into false GPU measurements consumed by operators.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-SANSAR-MECHElastic analyzed October activity against an unauthenticated Ray job-agent endpoint. The attacker inventoried NVIDIA hardware, introduced command execution and ran cryptomining components while deploying kernel and library-level concealment. The campaign used a replacement libnvidia-ml.so.1 to falsify process enumeration, load and temperatures, alongside a rootkit and Tor-capable implant. A separate on-chain Ethereum read supplied a changing controller address; Ethereum here served command-and-control discovery rather than proving wallet theft.supported

Basis: reported finding

Link to claim
CLM-NOV15-SANSAR-SCOPEThe report describes three compromised Linux GPU hosts and miner execution on two. It does not identify the initial root escalation method or establish mining revenue. Zero displayed GPU utilization is an attacker-controlled telemetry value, not evidence of idle hardware. The AI-security significance is a compromised shared accelerator or training environment that can defeat ordinary resource-health measurements while an attacker uses the same compute.supported

Basis: reported finding

Link to claim

Implications

The compromise crosses from job admission into host execution and then from host compromise into false GPU measurements consumed by operators.

Controls and mitigations

  • Authenticate and network-confine Ray job agents separately from the dashboard; restrict host rights for jobs.
  • Compare NVML user-space readings against independently collected device power, PCIe, kernel/module and process telemetry; investigate library replacement before trusting a clean utilization dashboard.

Unknowns and contradictions

  • The AI-security significance is a compromised shared accelerator or training environment that can defeat ordinary resource-health measurements while an attacker uses the same compute.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250