Security researchers have discovered that thousands of exposed Nvidia GPU monitoring systems are leaking sensitive telemetry data about AI infrastructure, with many also vulnerable to attacks that could crash the monitoring service. Lava Security published findings in a blog post showing that Nvidia's DCGM Exporter software contains an unauthenticated resource exhaustion flaw that enables remote attackers to disrupt monitoring and potentially interfere with AI workloads running on the same machine. Nvidia released a patch for the vulnerability in May after assigning it CVE-2026-47483 with a high severity rating of CVSS 8.2.
Lava researcher Michael Katchinskiy reported finding more than 2,000 GPU servers exposing Nvidia DCGM Exporter directly to the public internet without authentication. Over four separate scans, these servers revealed more than 12,000 distinct GPUs, representing an estimated $100 million worth of hardware. Approximately a quarter of the exposed instances were running profiling endpoints linked to the vulnerability. The exposed systems included Nvidia Blackwell Ultra B300 GPUs, H200s and H100s deployed for large-scale AI operations, plus consumer RTX 5090 and 4090 cards.
"Anyone who could reach these endpoints could see what hardware organizations were running, how heavily it was being used, and details about the AI infrastructure around it," Katchinskiy warned in the post. The weakness stems from how unpatched DCGM Exporter versions process simultaneous, unauthenticated requests to their "/debug/pprof/" profiling endpoints. A flood of concurrent profiling requests can push memory consumption high enough to deplete resources and crash the exporter. The telemetry agent collects metrics from Nvidia GPUs—including utilization, memory usage, power consumption, and error events—and exposes the metrics over HTTP for monitoring platforms such as Prometheus to scrape.
The immediate result is a loss of visibility into GPU health and activity, according to Lava, but the impact may extend beyond monitoring because the exporter shares a host with GPU workloads. The CPU and memory pressure could also disrupt AI training or inference processes operating alongside it. While the exposed information doesn't provide access to model weights, training data, or other protected assets, it can help an attacker profile a target and spot weaker components or software versions. Nvidia's security bulletin identifies DCGM Exporter version 4.8.2 as the updated release addressing the issue.
Lava recommends restricting public access to DCGM Exporter and Prometheus unless external reach is necessary, binding exporters to loopback or private interfaces, and enforcing access controls through firewalls, security groups, or other network measures. These steps should follow upgrading to version 4.8.2 or later and verifying that the "–enable–pprof" option is disabled. Katchinskiy noted that this profiling feature may be enabled by default in unpatched software. Organizations running GPU infrastructure for AI workloads should audit their exposure immediately and apply network-level controls alongside the software update to eliminate both the denial-of-service risk and the information leakage. The discovery underscores how operational tooling built for internal networks becomes a liability when deployed without authentication in cloud or hybrid environments. Enterprises accelerating AI adoption may overlook monitoring infrastructure in their threat models, creating blind spots that adversaries can exploit to map high-value compute assets before launching more targeted campaigns.