CVE-2026-47483

This CVE exposes the pprof profiling endpoint in NVIDIA's DCGM Exporter, a metrics collector for GPU-accelerated infrastructure. The vulnerability isn't simply a DoS risk from concurrent profiling requests—it's the exposure of a high-information-disclosure endpoint in a deployment context that elevates its severity beyond what CVSS 8.2 captures. The pprof endpoint reveals internal memory addresses, goroutine stacks, function call graphs, and potentially request contexts. In typical web services, this discloses deployment topology. In DCGM Exporter's environment—ML training pipelines, HPC clusters, enterprise AI deployments—it exposes the actual computational work product: model architectures, training data patterns, allocation fingerprints tied to specific model designs. This is competitive intelligence and espionage risk, not just operational disclosure. The privileged container context compounds the risk significantly. DCGM Exporter requires GPU device access, meaning it typically runs with elevated container capabilities. An attacker who accesses pprof gains profiling data *and* a foothold in a privileged container context—making this far more dangerous as a force multiplier in multi-stage intrusion than as a standalone entry point. For remediation, evaluate three options: binding pprof to localhost preserves diagnostic capabilities while blocking network access; adding authentication preserves remote debugging but introduces complexity; disabling pprof entirely removes the risk but loses debugging capability. Network-level controls (segmentation, ACLs) are a valid compensating control if code changes aren't immediately feasible, but they create a fragile security boundary that depends on infrastructure configuration rather than defense in depth. The low EPSS score (0.00307) should not minimize urgency—this endpoint is most dangerous as reconnaissance in an attack chain, particularly in environments where DCGM Exporter is deployed alongside less-hardened components. Prioritize remediation in any deployment where DCGM Exporter is network-accessible from non-trusted segments.

Reviewed through automated stages and approved by a human before publication.