Nvidia Patches DCGM Exporter Flaw Threatening AI Training Stability
Researchers found thousands of exposed GPU monitors vulnerable to resource exhaustion attacks that could disrupt AI training workloads on the same host systems.
Key points
- Nvidia released a patch in May for CVE-2026-47483, a high-severity flaw in DCGM Exporter.
- Lava Security researchers discovered over 2,000 unauthenticated instances exposed to the public internet.
- The vulnerability allows remote attackers to crash the service via concurrent requests to profiling endpoints.
Nvidia issued a security update in May to address a high-severity vulnerability in its DCGM Exporter software. The flaw allows unauthenticated remote attackers to exhaust system resources and crash the monitoring service.
What happened
Researchers at Lava Security identified CVE-2026-47483 as an uncontrolled resource-consumption vulnerability within the Nvidia DCGM Exporter telemetry agent. According to CSO Online, the component collects metrics from Nvidia GPUs for monitoring systems like Prometheus. The vulnerability exists in how the software handles concurrent, unauthenticated requests to its "/debug/pprof/" profiling endpoints. A large volume of these requests can drive memory consumption high enough to crash the exporter service. Nvidia assigned the flaw a CVSS score of 8.2, classifying it as high severity. The vendor released version 4.8.2 to address the issue.
Lava Security scanned the internet and found more than 2,000 GPU servers exposing DCGM Exporter without authentication. These servers reported over 12,000 unique GPUs across four scans. The exposed hardware represents an estimated $100 million in value, according to Lava researcher Michael Katchinskiy. About a quarter of the reachable instances exposed the specific profiling endpoints associated with the vulnerability. The researchers reported the finding to Nvidia before publishing their analysis.
Why it matters
The immediate impact of exploiting this flaw is a denial of service against the monitoring agent. This results in a loss of visibility into GPU health and activity for security operations teams. However, the consequences may extend beyond the monitoring layer because the exporter shares the host with active AI workloads. The CPU and memory pressure caused by the resource exhaustion can affect AI training or inference processes running alongside the exporter. This could lead to disruption of critical computational tasks on the same machine.
The exposure of profiling endpoints also facilitates information disclosure attacks. Remote attackers can access data without credentials, potentially revealing infrastructure secrets. The combination of unauthenticated access and resource exhaustion creates a significant risk for enterprises relying on continuous GPU monitoring. The widespread exposure of these services highlights the need for strict network segmentation and authentication controls for telemetry agents.
Our take on the response
Nvidia addressed the vulnerability by releasing version 4.8.2 of the DCGM Exporter. The update fixes the uncontrolled resource-consumption issue in the profiling endpoints. The vendor assigned CVE-2026-47483 to track the flaw publicly. The response came after Lava Security disclosed the widespread exposure of unauthenticated instances. The high CVSS score reflects the ease of exploitation and potential impact on shared host resources. Organizations should verify their installations are updated to mitigate the risk.
What to do and how to stay safe: Nvidia
- Audit network configurations to ensure DCGM Exporter instances are not directly accessible from the public internet.
- Implement authentication controls for all telemetry endpoints to prevent unauthenticated access to profiling data.
- Monitor system logs for unusual spikes in memory or CPU usage that may indicate resource exhaustion attempts.
- Update DCGM Exporter to version 4.8.2 or later once the vendor provides an update to address CVE-2026-47483.
Step-by-step guide: Multi-Factor Authentication Best Practices for Secure Access
General security guidance from the NetSec Tribune newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the CVSS score for CVE-2026-47483?
Nvidia assigned the vulnerability a CVSS score of 8.2, which is classified as high severity.
Which version of DCGM Exporter fixes the flaw?
Nvidia released version 4.8.2 to address the uncontrolled resource-consumption vulnerability.
How many exposed servers did researchers find?
Lava Security found more than 2,000 GPU servers exposing DCGM Exporter without authentication.
