> ## Content Index > Fetch the complete content index at: https://f4n6.co.uk/llms.txt > Use this file to discover other available public pages before exploring further. # High-severity Nvidia bug could crash GPU monitoring on exposed servers - URL: https://f4n6.co.uk/security-feed/high-severity-nvidia-bug-could-crash-gpu-monitoring-on-exposed-servers/ - Published: 2026-10-08T23:41:14.000Z - Updated: 2026-10-08T23:41:14.000Z - Author: Jeff Davies - Tags: #security-feed, CVE-2026-47483 ## 1\. Executive summary Nvidia's DCGM Exporter, a GPU telemetry service that publishes hardware, utilisation, memory, power and error metrics in plaintext over HTTP, carries a high-severity flaw tracked as CVE-2026-47483 (CVSS 8.2, CWE-770). Lava researcher Michael Katchinskiy found roughly 2,100 GPU servers exposing DCGM Exporter metrics to the open internet across four scans between March and May 2026, covering about 12,000 GPU UUIDs and roughly 300 organisations, none requiring authentication. The flaw is not in CISA KEV and EPSS sits at 1%, so mass opportunistic exploitation is not indicated; the exposure is the problem, because unauthenticated concurrent requests to the exporter can exhaust memory and crash it, cutting GPU health visibility and pressuring AI training or inference workloads. For EMEA financial services the direct risk is to firms running self-hosted GPU clusters for model training, risk analytics or inference, where an internet-reachable exporter is both a reconnaissance source and a denial-of-service target. Nvidia fixed the issue in DCGM Exporter 4.8.2; operators should be on that version or later. ## 2\. Regulatory framing | Article | Trigger (the fact in this item) | Practical impact | | ------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | DORA Art. 17: ICT-related incident management process | An unauthenticated internet-reachable monitoring endpoint can be crashed by concurrent requests, degrading GPU telemetry that feeds AI workload operations | Firms running DCGM Exporter must be able to detect and manage the loss of a monitoring service as an ICT-related incident, not treat it as a tooling glitch | | DORA Art. 28: ICT third-party risk — general principles | The exposures were reported across neocloud and GPU cloud providers including Nebius, Voltage Park, Lambda, Northern Data and DigitalOcean, so the endpoint may sit in a provider's estate rather than the client's | Where GPU capacity is consumed as a service, the client needs contractual visibility of the provider's monitoring-plane exposure and a route to require remediation | | NIS2 Art. 21(2)(d): supply chain security measures | GPU cloud and neocloud providers are suppliers in the AI compute chain, and the exposed service is part of their monitoring plane | Entities in scope must treat GPU compute suppliers as part of supply chain security measures, including which management and telemetry interfaces are internet-facing | No article from the regulatory reference is engaged by the CVE's severity alone. The triggers above rest on the specific fact that the exposed service is an unauthenticated monitoring endpoint inside a third-party GPU compute supply chain. ## 3\. Technical analysis & attack chain The vulnerability is CWE-770, allocation of resources without limits or throttling, in Nvidia's DCGM Exporter. The verified reference data records CVSS 8.2 HIGH and confirms the flaw is not in CISA KEV with an EPSS of 1%. The Register's reporting attributes an 8.2 CVSS rating to Nvidia's own advisory; the verified value matches, so there is no scoring discrepancy to reconcile. The source does not name a CWE, so no classification conflict arises. Confirmed steps, as described in the source: 1. DCGM Exporter runs on a GPU host and reads telemetry from the GPUs on that host: hardware identity, utilisation, memory usage, power consumption and error events. Each GPU has a unique UUID, and all metrics are served in plaintext over HTTP. 2. Where the exporter is reachable from the public internet, that telemetry is readable without authentication. Lava's scans between March and May 2026 found about 2,100 such hosts, exposing roughly 12,000 GPU UUIDs across about 300 organisations. 3. An attacker sends concurrent unauthenticated requests to the exporter. With enough concurrency, the exporter runs out of memory and crashes. 4. The crash removes visibility into GPU health and activity, and the CPU and memory pressure generated can affect AI training or inference workloads running on the same host. The reconnaissance value is separate from the crash. Exposed metrics let an attacker map GPU infrastructure, identify which systems are potentially vulnerable, and monitor workload activity. About 25 percent of the exposed DCGM hosts also served data from Go's `/debug/pprof/` profiling endpoint, which exposes CPU usage, memory allocations, goroutine states and blocking events for running Go applications. That endpoint widens the picture of the host's runtime behaviour and gives an attacker a second unauthenticated surface on the same service. The hardware involved is not marginal. Lava's scan data covers Nvidia Blackwell Ultra B300, H200 and H100 GPUs used for large-scale AI workloads, plus consumer RTX 5090 and 4090 systems, representing roughly $100 million in hardware. Nearly half of the exposed GPUs, 5,274 or 44 percent, were located in the US. The source gives no EMEA-specific count, so the regional distribution of exposed hosts is not established by this reporting. Lava also examined Prometheus Node Exporter, which exposes server hardware and OS metrics over HTTP. It found 12,096 public Node Exporter hosts publishing server models, operating systems, firmware versions, hostnames, storage paths and networking hardware typical of GPU clusters. That data describes how an environment is built and configured, which supports matching a host to known vulnerabilities. Node Exporter is not covered by CVE-2026-47483; it is a separate exposure finding in the same research. Confidence caveats. The scan counts, the 300-organisation figure, the 44 percent US share, the $100 million hardware valuation and the 25 percent pprof overlap all come from a single source, Lava's research as reported by The Register. Treat them as single-sourced; verify before using them in any internal risk quantification. The affected provider list (Nebius, Voltage Park, Lambda, Northern Data, DigitalOcean) is likewise single-sourced. The source states that Lava reported the exposures to those providers and that they worked with customers to address them; it does not state that all exposures are remediated. No exploitation in the wild is reported, and the verified data confirms CVE-2026-47483 is not in CISA KEV. No threat actor is named in the source, so there is no attribution to assess and no MITRE ATT&CK profile applies. ## 4\. Mitigation & containment P1 within 24h - Remove internet reachability from DCGM Exporter. Restrict it to authorised monitoring infrastructure, which is the vendor-independent control Lava recommends for DCGM Exporter, Node Exporter and Prometheus services. Where the exporter must be reachable across network segments, put it behind an authenticated reverse proxy or a service mesh with mTLS rather than exposing the HTTP listener directly. - Identify every host running DCGM Exporter and check its version. Upgrade to 4.8.2 or later, the version in which Nvidia fixed CVE-2026-47483. - Check for Go `/debug/pprof/` exposure on the same hosts. If the profiling endpoint is reachable, disable it or bind it to localhost; it is not needed for metric scraping. - If GPU capacity is consumed from a neocloud or GPU cloud provider, ask the provider in writing whether DCGM Exporter, Node Exporter or Prometheus endpoints in your tenancy are internet-reachable, and require confirmation of the version in use. P2 within 72h - Audit Prometheus Node Exporter instances for the same class of exposure. The source found 12,096 public hosts leaking server model, OS, firmware, hostname, storage path and networking hardware detail. Bind Node Exporter to the monitoring network and confirm no collector endpoint answers on a public interface. - Review firewall and security group rules for the exporter's listening port across all GPU subnets, including ephemeral cluster builds where a default rule may have been inherited from a template. - Add authentication and rate limiting at the monitoring ingress. Rate limiting directly addresses the CWE-770 mechanism by capping concurrent unauthenticated requests before they exhaust memory. P3 within 7 days - Inventory GPU UUIDs and host identities that were exposed, and treat the exposure window as a reconnaissance event. The leaked data describes infrastructure layout and workload activity; assess whether that changes the priority of any other open vulnerability on those hosts. - Add external attack surface monitoring for the exporter's HTTP response signature so a new internet-facing instance is detected on creation rather than at the next scan. - Where the GPU estate sits with a third-party provider, fold monitoring-plane exposure into the supplier review cycle and record the provider's answer from P1. ## 5\. Indicators of compromise No indicators of compromise available in the source material. The source describes exposed services and a crash mechanism, not attacker infrastructure, files or network artefacts. No hashes, domains, IP addresses or file paths are given. Behavioural indicators | behaviour | where to observe | confidence | | ------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------ | | Unauthenticated HTTP requests to the DCGM Exporter metrics endpoint from outside the monitoring network | Exporter access logs, reverse proxy logs, network flow records on the GPU subnet | High for the exposure pattern; the source does not publish a request signature | | Concurrent request bursts against the exporter followed by process termination or OOM kill | Host system logs, container runtime events, exporter process exit records | High; the crash mechanism is described in the source | | Requests to Go /debug/pprof/ paths on GPU hosts | Web server or exporter access logs | Medium; the source reports pprof exposure on about 25 percent of scanned hosts but gives no request examples | | Scraping of Node Exporter metrics from public sources | Node Exporter access logs, edge firewall logs | Medium; separate exposure finding, not tied to CVE-2026-47483 | ## 6\. Detection The sources give no file hashes, mutexes, registry keys or command-line artefacts. What they do give is a distinctive HTTP path structure for the exposed services, which is enough for a network or web-log detection rule. No YARA rule is emitted: there is no malicious file in this item to match against, and a rule built from the CVE identifier or the product name would detect reporting about the flaw rather than the flaw itself. ```yaml title: Unauthenticated Access to GPU Monitoring Exporter Endpoints id: 7f3c1a92-4d5e-4b8a-9c21-0e6a2b7d4f18 status: experimental description: Detects HTTP requests to Nvidia DCGM Exporter metrics endpoints and Go pprof profiling paths from outside the authorised monitoring network. DCGM Exporter serves GPU telemetry in plaintext over HTTP without authentication; CVE-2026-47483 allows concurrent unauthenticated requests to exhaust memory and crash the exporter. author: Adverse Trace date: 2026-10-09 references: - https://www.theregister.com/security/2026/10/08/high-severity-nvidia-bug-could-crash-gpu-monitoring-on-exposed-servers/5302077 logsource: category: webserver detection: selection_dcgm: cs-uri-stem|startswith: - '/metrics' selection_pprof: cs-uri-stem|contains: - '/debug/pprof/' filter_internal: src_ip|cidr: - '10.0.0.0/8' - '172.16.0.0/12' - '192.168.0.0/16' condition: (selection_dcgm or selection_pprof) and not filter_internal falsepositives: - Legitimate monitoring infrastructure that scrapes from outside the listed RFC1918 ranges; tune filter_internal to the organisation's collector addresses level: medium ``` The `/metrics` path is the default Prometheus scrape path used by DCGM Exporter and Node Exporter, and `/debug/pprof/` is Go's built-in profiling prefix; both appear in the source. Tune the internal-range filter to your own collector addresses before enabling, or the rule will fire on normal scraping. ## CVE assessment 1 referenced CVE | CVE | CVSS | Exploited | EPSS | Summary | | -------------------------------------------------------------------------------- | -------- | --------- | ---- | -------------------------------------------------------------------------------------------------------------------------------- | | [CVE-2026-47483](https://nvd.nist.gov/vuln/detail/CVE-2026-47483?ref=f4n6.co.uk) | 8.2 High | — | 1% | NVIDIA DCGM Exporter for all platforms contains a vulnerability in the /debug/pprof endpoints, where an attacker could cause un… | ## 7\. Sources - The Register, "High-severity Nvidia bug could crash GPU monitoring on exposed servers", https://www.theregister.com/security/2026/10/08/high-severity-nvidia-bug-could-crash-gpu-monitoring-on-exposed-servers/5302077, 2026-10-08 - BSI Germany, "\[UPDATE\] \[hoch\] Nvidia GPU Display Treiber: Mehrere Schwachstellen", https://wid.cert-bund.de/portal/wid/securityadvisory?name=WID-SEC-2025-2247, undated in the supplied extract. This advisory covers Nvidia GPU display driver vulnerabilities and is not the DCGM Exporter flaw; it is listed for completeness of the supplied material and is not relied on for any claim above. ## 8\. Adverse Trace position Adverse Trace assesses CVE-2026-47483 as CVSS 8.2 HIGH, not in CISA KEV, with EPSS at 1 percent, so we do not expect opportunistic mass exploitation and we are not raising it to a campaign-level alert. The material risk for EMEA financial services clients is the exposure pattern rather than the CVE: an unauthenticated HTTP monitoring endpoint on a GPU host that both describes the AI estate to anyone who asks and can be crashed by concurrent requests. Clients running self-hosted GPU clusters for training, risk analytics or inference should treat internet-reachable DCGM Exporter, Node Exporter and Prometheus endpoints as a P1 finding today, upgrade DCGM Exporter to 4.8.2 or later, and confirm with any neocloud or GPU cloud provider whether their monitoring plane is exposed in the client's tenancy. The scan statistics in this advisory are single-sourced to Lava via The Register; we will seek corroboration from Nvidia's own advisory and from the named providers before we publish exposure figures to clients as established fact. We will also monitor for the first credible report of exploitation, which would change the KEV status and our assessment. --- [Read the original source →](https://www.theregister.com/security/2026/10/08/high-severity-nvidia-bug-could-crash-gpu-monitoring-on-exposed-servers/5302077?ref=f4n6.co.uk) *Published via PulseTrace — Adverse Trace threat intelligence.*