NVIDIA DCGM Exporter flaw CVE-2026-47483 can crash GPU monitoring on exposed hosts

Researchers at datacenter security company Lava published details on 8 October 2026 of CVE-2026-47483, a flaw in NVIDIA’s DCGM Exporter, the service that exposes GPU telemetry to Prometheus. NVIDIA’s security bulletin, dated 28 July 2026, credits Lava researcher Michael Katchinskiy. This is a catch-up: the fix has been out for weeks, but the disclosure and the exposure data are new.

The bug

NVIDIA’s description: DCGM Exporter “contains a vulnerability in the /debug/pprof endpoints, where an attacker could cause uncontrolled resource consumption by submitting concurrent unauthenticated profiling requests.” A successful exploit “might lead to denial of service and information disclosure.” NVIDIA scores it CVSS 3.1 8.2, High, network-reachable with no privileges, classed as CWE-770.

pprof is Go’s built-in profiler. Lava says some profiling endpoints hold a request open for a caller-chosen duration, so enough concurrent requests drive memory up until the exporter crashes. That blinds operators to GPU health, and Lava says the CPU and memory pressure can slow training or inference on the same host. Lava reproduced it with NVIDIA’s official container, unmodified, and tested the exhaustion in a controlled environment, not against public hosts.

Which version fixes it

NVIDIA’s bulletin decides this. Its table lists:

Product	Affected	Updated version
DCGM	up to 4.5.2	4.5.3
DCGM Exporter	up to 4.8.2	4.8.2

The exporter row lists 4.8.2 in both columns, which is why some coverage points to 4.8.3 instead. The Register reports 4.8.2 as the fix. The current exporter release, 4.8.3 with DCGM 4.6.0 (15 July), lists “Strengthen opt-in pprof and socket validation” among its changes. Running 4.8.3 satisfies either reading.

How exposed are GPU fleets

These are Lava’s figures. Across four internet scans between March and May 2026, it saw roughly 2,100 hosts publicly serving DCGM Exporter metrics over plain HTTP with no authentication, reporting more than 12,000 unique GPU IDs across nearly 300 organisations. About a quarter of those hosts also served /debug/pprof. Lava estimates the exposed hardware at $100 million, and says it included 312 B200s and 32 B300s alongside H100s, H200s and consumer RTX 4090 and 5090 cards. By location, Lava counted 44% of the exposed GPUs in the United States, 17% in Romania and 16% in China. Many sat on customer infrastructure at GPU cloud providers, whose security teams, Lava says, helped notify customers.

Even without the CVE, Lava notes, exposed metrics reveal GPU models, utilisation, schedules and Kubernetes pod names.

What to do

• Upgrade DCGM Exporter to 4.8.2 or later (4.8.3 is current) and DCGM to 4.5.3 or later.
• Bind the metrics port (9400 by default) to a private interface or put it behind authentication; Prometheus does not need it public.
• Make sure /debug/pprof is not reachable from outside the host.