The Bottom Line

Researchers found a high-severity flaw in NVIDIA DCGM Exporter, tracked as CVE-2026-47483 with a CVSS score of 8.2. The bug lets unauthenticated attackers exhaust server resources and crash GPU monitoring remotely.

The same research effort found more than 2,100 exposed GPU monitoring servers sitting on the open internet. Those servers exposed telemetry from over 12,000 GPUs, including H100s, H200s, and Blackwell units, with an estimated hardware value of roughly $100M.

NVIDIA has fixed the vulnerability following responsible disclosure from the research team. Organizations running DCGM Exporter should confirm their deployment status and check for public exposure immediately.

What Happened

NVIDIA DCGM Exporter is a monitoring tool widely deployed across AI infrastructure. It collects GPU metrics such as utilization, memory usage, temperature, and power consumption from servers running NVIDIA hardware.

Security researchers scanning the public internet discovered thousands of these monitoring endpoints left exposed without authentication. Anyone could query them and inspect the underlying GPU infrastructure in detail.

During that research, the team identified a deeper problem beyond simple exposure. The monitoring service itself contained a vulnerability that let unauthenticated attackers send requests capable of exhausting server resources.

That resource exhaustion can crash the GPU monitoring process outright. Because monitoring often sits close to orchestration and scheduling logic in AI environments, the researchers warned the flaw could potentially disrupt AI workloads remotely.

The researchers disclosed the vulnerability to NVIDIA through a responsible disclosure process. NVIDIA fixed the issue in response, and the fix is now reflected in the assigned identifier, CVE-2026-47483.

The team published full technical details of the vulnerability and their internet-wide scanning methodology in a public research report. The report also documents the scale of exposed GPU monitoring infrastructure they found during the same effort.

Who Is Affected

Any organization running NVIDIA DCGM Exporter with an internet-facing endpoint is in scope for this issue. That includes cloud providers, AI infrastructure operators, and enterprises running on-prem GPU clusters for training or inference.

The research identified 2,100-plus exposed GPU servers running the monitoring tool. Those servers revealed telemetry tied to more than 12,000 individual GPUs, spanning NVIDIA's current high-end lineup.

Exposed GPU models included H100s, H200s, and Blackwell-generation chips, the hardware class that underpins most large-scale AI training today. A separate set of 12,096 servers exposed additional infrastructure details beyond GPU telemetry.

That additional exposure included operating system data, firmware versions, and networking information. Combined, these two exposure categories give attackers a detailed map of target AI infrastructure before launching any attack.

By the Numbers

CVE-2026-47483 carries a CVSS score of 8.2, rated high severity. Researchers found 2,100-plus exposed GPU servers running DCGM Exporter publicly accessible.

Those servers exposed telemetry from 12,000-plus individual GPUs. The research team estimated the value of that exposed hardware at approximately $100M.

Sixty percent of exposed GPUs reported 0% utilization across the research team's scans. A separate group of 12,096 servers exposed additional infrastructure details, including OS, firmware, and networking information.

NVIDIA fixed the vulnerability following the research team's responsible disclosure.

Why It Matters

This case shows a pattern familiar to anyone securing AI infrastructure: monitoring tools often ship with weaker default security postures than the workloads they observe. Teams harden model endpoints and training clusters, then leave telemetry dashboards open by default.

The exposed utilization data tells its own story. Sixty percent of discovered GPUs sat idle during scans, meaning organizations are leaking capacity and spending data to anyone scanning the internet.

Combine that exposure with a remotely exploitable denial-of-service flaw and the risk compounds quickly. An attacker does not need credentials to crash monitoring and potentially disrupt the AI workloads that monitoring was built to protect.

Given the value concentrated in modern GPU fleets, roughly $100M in hardware exposed in this research alone, the stakes for misconfigured monitoring are higher than in traditional IT environments.

What To Do Now

First, inventory every instance of NVIDIA DCGM Exporter running in your environment, cloud or on-prem. Second, confirm whether any instance is reachable from the public internet and remove that exposure immediately.

Third, apply NVIDIA's fix for CVE-2026-47483 across all affected DCGM Exporter deployments without delay. Fourth, place remaining monitoring endpoints behind authentication and network segmentation rather than relying on obscurity.

Fifth, review logs for unusual or high-volume requests to DCGM Exporter endpoints that could indicate scanning or resource-exhaustion attempts. Sixth, treat exposed firmware, OS, and networking details as reconnaissance data and rotate any credentials or configurations that may have been visible.

Finally, read the full technical research report before completing remediation, since it documents the exact request patterns used to trigger resource exhaustion.

Read the source →