Nvidia Monitoring Bug Could Have Knocked AI Offline

Nvidia Monitoring Bug Could Have Knocked AI Offline 🚨

A recent Nvidia monitoring bug could have potentially knocked AI systems offline. This vulnerability allowed attackers to crash monitoring services and disrupt AI workloads without any authentication. Researchers discovered that over 2,000 GPU servers were exposing Nvidia’s DCGM Exporter online without proper security measures.

Key Findings:

• The flaw enabled unauthenticated attackers to crash the monitoring service, risking disruption to AI workloads.
• Exposed systems included around 300 organizations, representing approximately $100 million in hardware.
• Nvidia has confirmed the issue and has released a fix for this high-severity flaw.

The DCGM Exporter tool reads telemetry from GPUs, including hardware information, memory usage, power consumption, and error events. Researchers from Lava, a cybersecurity company, highlighted that this bug could lead to denial of service and information disclosure by exhausting the exporter’s resources.

Impact:

The exposed GPUs included Nvidia’s Blackwell Ultra B300, H200s, and H100s, which are crucial for large-scale AI workloads, as well as consumer models like the RTX 5090 and 4090. Anyone with access to these endpoints could view detailed information about the hardware and the AI infrastructure.

Yakir Kadkoda, CTO and Co-Founder at Lava, stated, “Neoclouds are racing to add GPU capacity, and customers are racing to use it. That speed is creating security gaps on both sides.”

Lava reported the issue to Nvidia, which acknowledged the problem and released a fix tracked as CVE-2026-47483, rated with an 8.2 CVSS high-severity rating. This incident underscores a growing security gap in AI infrastructure, where companies invest heavily in GPUs while leaving critical systems exposed.

For more details, check the full article here: Read full article