Censys internet census finds sharp growth in externally visible AI infrastructure

The Censys 2026 State of the Internet public summary reports 63 percent growth in exposed AI infrastructure over a nine-month comparison. It separately reports that 76 percent of n8n hosts whose version could be identified are affected by a critical vulnerability. These are exposure-inventory metrics from a provider's scanning methodology and not counts of hosts proven accessible without authentication or exploited by malicious actors.

The security boundary is material for agent infrastructure because public execution and automation control surfaces can expose high-impact tooling, but reachability depends on specific service configurations and credentials. The published top-line summary does not disclose the full host denominator, version-identification bias, deduplication method or per-product validation here. More granular counts repeated in press summaries remain unverified until the full original report and methodology can be audited.

The denominator problem applies twice: the nine-month infrastructure comparison depends on stable service signatures across observation dates, while the n8n vulnerability figure is limited to hosts for which a software version was identified. Neither group is automatically the population of remotely exploitable agent installations. Even when a host presents an automation interface, credentials, network policy and application configuration determine whether an attacker can operate consequential tools. Treat inventory growth, vulnerable-version detection and unauthorized execution as three separate measurements.