NVIDIA patched high-severity flaw CVE-2026-47483 in DCGM Exporter that allows unauthenticated remote attackers to exhaust server resources, crash GPU monitoring, and disrupt AI workloads.
DCGM Exporter is a common monitoring tool for GPU servers and Kubernetes clusters. It reads details such as GPU model, temperature, memory use, power draw, errors, and workload activity, then serves the data in a Prometheus-friendly format over HTTP.
A unique UUID also lists each GPU. If this endpoint is public, an outsider can map valuable hardware, watch busy periods, and learn details about the software and workloads around it.
The weakness sits in Go profiling routes under /debug/pprof. These routes were meant to help developers inspect CPU and memory use. However, vulnerable builds exposed them without login checks or safe request limits.
By sending many profiling requests at once, an attacker could cause uncontrolled memory and CPU use until the exporter slowed, stopped responding, or crashed.
NVIDIA DCGM Exporter Bug Exposes
NVIDIA rated the issue 8.2 out of 10 and classified it as CWE-770, which covers resource allocation without proper limits. A crash would remove the live view of GPU health, errors, temperature, and use.
The same resource pressure could also slow nearby AI tasks when the exporter lacks strict container or host limits. The researchers tested the resource exhaustion only in a controlled environment and did not target the public systems they found.
Lava Security researcher Michael Katchinskiy found the flaw while studying public DCGM Exporter systems. Four internet scans between March and May 2026 identified about 2,100 hosts serving metrics without authentication.
They exposed more than 12,000 unique GPU identifiers across about 300 organizations. Around one quarter of the DCGM hosts also exposed the risky profiling routes.
The visible hardware included H100, H200, B200, Blackwell Ultra B300, RTX 4090, and RTX 5090 GPUs. The United States accounted for 5,274 exposed GPUs, or 44 percent of the observed total.
Researchers also found customer-run endpoints on networks linked to GPU cloud providers, including Voltage Park, Lambda, Northern Data, DigitalOcean, and others. The providers helped alert customers after receiving the findings.
The case shows why AI monitoring ports should receive the same care as other management services. Cyber Security News has previously covered NVIDIA GPU driver flaws that enable code execution and published an AI data center security checklist covering GPU, firmware, and management-layer risks.
NVIDIA’s security bulletin directs DCGM Exporter users to install version 4.8.2 or later. Operators should also keep –enable-pprof disabled unless profiling is needed.
Exporter, Prometheus, and Node Exporter services should bind to a private or loopback interface and accept traffic only from approved monitoring systems through firewall rules, cloud security groups, or Kubernetes NetworkPolicies.
Security teams should review internet-facing port 9400, block public access to /debug/pprof, and inspect logs for repeated or long-running profiling requests.
Sudden memory growth, CPU spikes, Prometheus scrape failures, or gaps in GPU alerts may signal probing or exploitation. Teams should confirm that patched containers are deployed across every GPU node, not merely updated in a registry or configuration file.
Stops threats before impact with 21 min faster MTTR. Integrate ANYRUN’s Sandbox in your SOC
