NVIDIA patched a high-severity vulnerability in DCGM Exporter caused by unprotected Go profiling routes.
The flaw allowed remote attackers to exhaust resources and disrupt AI workloads, highlighting configuration risks in GPU monitoring infrastructure.
The tool, known as DCGM Exporter, collects metrics from graphics cards and makes that information available online. However, in some cases, diagnostic functions intended for technical checks remained accessible without requiring a password or having a limit on requests.
According to the researchers who discovered the issue, Michael Kaczynski of Lava Security, the absence of these checks meant that someone could send mass requests and exhaust the computer's resources. This could cause graphics card monitoring to stop and affect artificial intelligence tasks.
Online searches conducted between March and May 2026 showed that there were 2.100 exposed systems across approximately 300 organizations, revealing unique identifiers for over 12.000 graphics cards. About a quarter of these systems had the dangerous diagnostics turned on.
The list of exposed equipment included powerful graphics card models such as the H100, H200, B200, Blackwell Ultra B300, RTX 4090, and RTX 5090. The United States collected 5.274 graphics cards, a number corresponding to 44 percent of the total, while the exposed points were found in the networks of cloud service providers such as Voltage Park, Lambda, Northern Data, and DigitalOcean.
NVIDIA rated the issue 8,2 out of 10 and urged administrators to install version 4.8.2 or later, while ensuring that specific diagnostic features remain disabled unless absolutely necessary.
Top comments (0)