Several hundred GPU servers sitting on the public internet were reachable to a high-severity flaw in NVIDIA’s DCGM Exporter, a monitoring tool that reads straight from the hardware. Tracked as CVE-2026-47483 with an 8.2 CVSS rating, the bug lets unauthenticated attackers crash the service.
The security firm Lava disclosed it to NVIDIA, and the vendor’s bulletin followed on July 28. The exporter surfaces temperature, utilization, memory and power figures over HTTP, typically on port 9400, for collectors like Prometheus to scrape.
Between March and May the team ran four passes, and the machines they reached reported more than 12,000 separate GPUs worth an estimated $100 million; not one host required authentication, and every metric moved over plaintext HTTP.
What the scans surfaced ranged from consumer RTX 5090 and 4090 cards on one end to the Blackwell Ultra B300, H200 and H100 accelerators NVIDIA designs for heavy AI work on the other. The hosts belonged to nearly 300 organizations; the United States accounted for 44 percent of the GPUs, ahead of Romania and China.
Go’s /debug/pprof/ profiling endpoints were also live on about a quarter of the exposed hosts. A flood of concurrent unauthenticated requests can drain memory and crash the exporter, cutting visibility into GPU health and slowing training or inference on the same machine. The team reproduced the effect on an unmodified official container.
The remedies are simple: upgrade DCGM Exporter to 4.8.2 or later, leave the profiling flag off, and bind monitoring services to loopback or private interfaces behind firewall rules.
