The service that tells your team whether your AI GPUs are healthy is often published over plain HTTP with no password. A high-severity flaw in Nvidia’s DCGM Exporter lets anyone take that monitoring down, and Lava research found about 2,100 GPU servers reachable from the internet.
Nvidia’s DCGM Exporter collects GPU metrics (utilization, memory, power draw, errors) and exposes them so tools like Prometheus can consume them. According to The Register, that data is published over unencrypted HTTP with no authentication.
The vulnerability, CVE-2026-47483 (CVSS 8.2, high severity; CWE-770), lives in the /debug/pprof debugging endpoints. According to vulnerability databases, an unauthenticated attacker can send many concurrent profiling requests and exhaust memory, CPU and file descriptors until the exporter crashes. The advisory is listed as published on July 28, 2026; Nvidia fixed the problem in version 4.8.2 or later.
Michael Katchinskiy, of the data center security startup Lava, reported the flaw. His research found about 2,100 GPU servers reachable from the internet and exposing unauthenticated metrics (figure reported by The Register and Diario Bitcoin). According to The Register, citing scans from March to May, that equates to roughly 12,000 GPU identifiers across about 300 organizations, 44% of them in the US, with models such as H100, H200, B300 and RTX 4090/5090 and an estimated hardware value of around US$100 million. These last figures come from that single source, so treat them as estimates.
Knocking out the exporter doesn’t shut down the GPUs, but it leaves the team blind: without metrics you can’t detect memory faults, overheating or abnormal consumption, nor unauthorized compute use such as mining or someone else’s training jobs. In an AI cluster where each GPU costs tens of thousands of dollars, losing visibility is a real operational risk. On top of that, the exposed metrics already reveal what hardware you have and how heavily you use it, which is useful to an attacker choosing targets.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel