NVIDIA DCGM Exporter Bug Exposes GPU Servers to Resource Exhaustion Attacks

Key Takeaways

• NVIDIA has addressed a high-severity vulnerability (CVE-2026-47483) in its DCGM Exporter tool.
• The flaw could allow unauthenticated remote attackers to trigger resource exhaustion, disrupting GPU monitoring and potentially impacting AI workloads.
• A significant number of DCGM Exporter instances, including those linked to major GPU cloud providers, were found publicly exposed without authentication.
• The vulnerability has a CVSS score of 8.2 and is categorized as CWE-770 (Improper Restriction of Resource Allocation).
• Users should update to DCGM Exporter version 4.8.2 or later and implement strict network access controls.

NVIDIA has released a critical patch for a high-severity vulnerability, identified as CVE-2026-47483, within its DCGM Exporter utility. This flaw could enable unauthenticated remote attackers to initiate resource exhaustion attacks, leading to the disruption of GPU monitoring services and potentially impacting critical AI workloads.

DCGM Exporter is an essential component for monitoring GPU servers and Kubernetes clusters. It gathers vital statistics such as GPU model, temperature, memory utilization, power consumption, error logs, and workload activity. This data is then made available in a Prometheus-compatible format via HTTP endpoints.

Each GPU is also assigned a unique UUID, which, if exposed, can provide external parties with valuable insights into an organization’s hardware infrastructure, peak usage periods, and details about the software and workloads running on these systems.

Vulnerability Details

The core of the vulnerability resides in the Go profiling routes, specifically under the /debug/pprof endpoint. These routes are typically intended for developers to analyze CPU and memory usage within the application. However, affected versions of DCGM Exporter exposed these profiling endpoints without requiring authentication or implementing adequate request limits, creating a significant security loophole.

Attackers could exploit this by sending a high volume of concurrent profiling requests. This barrage of requests would lead to uncontrolled consumption of memory and CPU resources, eventually causing the exporter to slow down, become unresponsive, or crash entirely. NVIDIA has assigned the vulnerability a CVSS score of 8.2 out of 10 and classified it under CWE-770, which addresses improper resource allocation without sufficient limits.

A crash of the DCGM Exporter would result in the loss of real-time visibility into GPU health, error states, temperature, and overall utilization. Furthermore, if the exporter operates without strict container or host resource limits, the generated resource pressure could also degrade the performance of co-located AI tasks. Researchers confirmed the resource exhaustion in controlled test environments, refraining from targeting live public systems.

Widespread Exposure Uncovered

The vulnerability was discovered by Lava Security researcher Michael Katchinskiy during an investigation into publicly accessible DCGM Exporter instances. Between March and May 2026, four internet scans revealed approximately 2,100 hosts that were serving metrics without any authentication requirements.

These exposed instances collectively revealed over 12,000 unique GPU identifiers across nearly 300 different organizations. Alarmingly, approximately a quarter of these DCGM hosts also had the vulnerable profiling routes openly exposed. The observed hardware included high-end accelerators like H100, H200, B200, Blackwell Ultra B300, as well as consumer-grade GPUs such as RTX 4090 and RTX 5090. The United States accounted for the largest share, with 5,274 exposed GPUs, representing 44 percent of the total.

The research also identified customer-managed endpoints on networks associated with prominent GPU cloud providers, including Voltage Park, Lambda, Northern Data, and DigitalOcean, among others. These providers were subsequently notified of the findings and assisted in alerting their respective customers.

This incident underscores the critical importance of treating AI monitoring ports with the same rigorous security protocols applied to other management services. Prior cybersecurity reports have highlighted NVIDIA GPU driver vulnerabilities leading to code execution, emphasizing the need for comprehensive security checklists encompassing GPU, firmware, and management-layer risks within AI data centers.

What You Should Do

• Update Immediately: Install DCGM Exporter version 4.8.2 or later as directed by NVIDIA’s security bulletin.
• Disable Profiling Routes: Keep the --enable-pprof option disabled unless profiling is absolutely essential for debugging or development purposes.
• Restrict Network Access: Configure Exporter, Prometheus, and Node Exporter services to bind only to private or loopback interfaces. Traffic should only be accepted from authorized monitoring systems via strict firewall rules, cloud security groups, or Kubernetes NetworkPolicies.
• Block Public Access to /debug/pprof: Review internet-facing port 9400 and ensure public access to the /debug/pprof endpoint is explicitly blocked.
• Monitor for Anomalies: Regularly inspect logs for any signs of repeated or unusually long-running profiling requests. Be alert for sudden memory growth, CPU spikes, Prometheus scrape failures, or gaps in GPU alerts, as these could indicate probing or active exploitation.
• Verify Deployment: Confirm that patched containers are deployed across all GPU nodes, rather than merely updating them in a registry or configuration file.

Disclaimer: HackersRadar reports on cybersecurity threats and incidents for informational and awareness purposes only. We do not engage in hacking activities, data exfiltration, or the hosting or distribution of stolen or leaked information. All content is based on publicly available sources.