NVIDIA monitoring tool flaw left thousands of GPUs exposed to denial-of-service attacks
NVIDIA's DCGM Exporter contained a high-severity bug, CVE-2026-47483, that could be crashed by attackers who did not need credentials, potentially disrupting AI workloads. Lava discovered more than 2,000 internet-facing servers during scans between March and May 2026, covering over 12,000 GPUs across nearly 300 organizations. NVIDIA gave the issue a CVSS score of 8.2 and released a security bulletin on July 28, 2026.
NVIDIA’s DCGM Exporter collects GPU telemetry—temperature, utilization, memory, power, and errors—and serves it over HTTP, usually port 9400, for Prometheus. Lava’s Michael Katchinskiy found over 2,000 internet-facing instances between March and May 2026, representing more than 12,000 GPUs across nearly 300 organizations. The flaw, CVE-2026-47483, received CVSS 8.2; NVIDIA issued a bulletin on July 28, 2026.
Roughly a quarter of exposed hosts also published Go /debug/pprof/ endpoints. Many simultaneous requests without credentials could exhaust memory and crash the exporter. Researchers also found 12,096 Node Exporter instances reachable publicly, revealing adapter and firmware identifiers, link status, host names, operating system, kernel, and BIOS information.
If such monitoring endpoints remain exposed, outages could blind operators to GPU health, potentially slowing or interrupting AI training and inference. Organizations running shared GPU clouds, research labs, and enterprises may face service disruptions, troubleshooting delays, and higher operational costs. Because the flaw needs no credentials, opportunistic actors could trigger crashes, though the reported issue affects availability rather than direct data theft. The broader lesson may be that observability services need the same hardening as production systems.