AI Infrastructure Vulnerability Discovered in Nvidia GPU Monitoring Tools
An unauthenticated resource exhaustion vulnerability in the Nvidia DCGM Exporter tool paves the way for attacks targeting AI infrastructures.
A critical flaw has been identified in a GPU monitoring software component used in Nvidia's AI infrastructures, which can lead to denial of service and information disclosure. Experts emphasize that thousands of servers are exposed to the internet and necessary updates must be applied.
Details and Risks of the Vulnerability
The unauthenticated resource exhaustion vulnerability detected in the Nvidia DCGM Exporter software paves the way for remote attacks. Tracked under the code CVE-2026-47483, this high-severity flaw can cause monitoring services to crash.
Research Findings
Scans conducted by Lava Security researchers showed that thousands of GPU servers host the Nvidia DCGM Exporter directly exposed to the internet. It is estimated that these servers contain hardware worth millions of dollars.
Mechanism of the Vulnerability
The problem stems from how unpatched DCGM Exporter versions handle concurrent, unauthenticated requests to profiling endpoints. Excessive memory consumption depletes system resources, causing services to crash.
Precautions to Take
While Nvidia has released updates addressing the issue, experts recommend restricting public access. Updating the software and enhancing network security measures are of great importance for security.