Lava Security Reveals NVIDIA Flaw Letting Attackers Crash $100M in GPUs Without Credentials

Lava Security researchers discovered a high-severity vulnerability in NVIDIA’s DCGM Exporter that allows unauthenticated attackers to disrupt GPU monitoring services remotely. The flaw, tracked as CVE-2026-47483, enables resource exhaustion and service crashes on internet-exposed servers, potentially impacting AI workloads.

What happened

Lava Security identified that thousands of instances of NVIDIA DCGM Exporter are exposed to the public internet, revealing infrastructure telemetry to anyone who scans them. According to the research team, these monitoring endpoints collect metrics such as utilization, memory, temperature, and power consumption from AI servers. The team found that 2,100 servers were accessible, representing telemetry from more than 12,000 GPUs.

The vulnerability allows unauthenticated attackers to send specific requests that exhaust server resources and crash the GPU monitoring service. This disruption can indirectly affect AI workloads running on the hardware. Lava Security reported the issue to NVIDIA, which rated the flaw 8.2 on the CVSS scale, classifying it as high severity. The vendor published a security bulletin addressing the issue on July 28, 2026.

Scans conducted by Lava Security revealed that 60% of the exposed GPUs reported zero utilization during the testing period. This indicates that many of these systems may be idle or reserved for future tasks. Despite low utilization, the exposure remains significant because the monitoring service is directly accessible from the internet without authentication requirements in many configurations.

The research also uncovered that 12,096 additional servers exposed infrastructure details beyond just GPU metrics. These details included operating system versions, firmware information, and networking configurations. Such data leakage provides attackers with a broader view of the target environment, facilitating further reconnaissance for potential exploits.

Why it matters

The exposure of 12,000 GPUs, including high-end models like the H100, H200, and Blackwell series, represents a substantial financial risk. The estimated value of this exposed hardware is approximately $100 million. For organizations relying on these systems for AI training or inference, service disruption could lead to significant downtime and loss of productivity.

The unauthenticated nature of the attack vector means that any individual with internet access can attempt to exploit the vulnerability. There is no requirement for stolen credentials or prior network access. This lowers the barrier for entry for malicious actors, increasing the likelihood of widespread automated scanning and exploitation attempts.

Furthermore, the leakage of infrastructure details such as OS and firmware versions creates secondary risks. Attackers can use this information to identify other known vulnerabilities in the server environment. The combination of a direct crash vulnerability and information disclosure amplifies the overall threat to the organization's security posture.

Our take on the response

NVIDIA’s rapid publication of a security bulletin following responsible disclosure demonstrates a standard response to high-severity findings. The CVSS score of 8.2 correctly reflects the high impact of unauthenticated remote denial of service. Organizations should prioritize patching DCGM Exporter instances immediately to mitigate the risk of service disruption.

What to do and how to stay safe: NVIDIA

• Audit firewall rules to ensure DCGM Exporter ports are not accessible from the public internet.
• Implement authentication mechanisms for all GPU monitoring endpoints to prevent unauthenticated access.
• Verify that all DCGM Exporter instances are updated to the latest patched version provided by NVIDIA.
• Monitor server logs for unusual resource exhaustion patterns that may indicate exploitation attempts.

Step-by-step guide: Vulnerability Management Myths That Waste Your Time and Budget

General security guidance from the NetSec Tribune newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is CVE-2026-47483?

CVE-2026-47483 is a high-severity vulnerability in NVIDIA DCGM Exporter that allows unauthenticated attackers to crash the GPU monitoring service and exhaust server resources.

How many GPUs were exposed in this study?

The research identified telemetry from more than 12,000 GPUs across 2,100 exposed servers, including H100, H200, and Blackwell models.

Did NVIDIA fix the vulnerability?

Yes, NVIDIA published a security bulletin on July 28, 2026, after Lava Security responsibly disclosed the flaw and provided proof of concept.