Summary
- Censys reports a 63% increase in exposed AI infrastructure over nine months.
- Its survey identifies roughly 134,000 internet-exposed industrial control hosts.
- An exposed service is not evidence that it has been breached.
Internet intelligence company Censys says the number of publicly exposed artificial intelligence infrastructure services grew 63% in nine months, while around 134,000 hosts associated with industrial control systems remained reachable from the internet. The findings appear in the company’s State of the Internet report released on 14 October 2026.
The two measurements describe different infrastructure groups. AI services include systems brought online as organisations experiment with models and deploy software around them. Industrial control hosts represent equipment or services tied to physical operations. Both can appear in internet scans without their owners intending them to be broadly accessible, although reachability alone does not establish that a system has been compromised or is exploitable.
Censys reports that approximately 70% of the industrial-control hosts it observed were on consumer or mobile networks. This complicates assumptions that every visible industrial service belongs to a directly identifiable operator with an established corporate security team. Equipment may be deployed through integrators, connectivity providers or remote maintenance arrangements, and address ownership does not necessarily identify operational ownership.
An inventory built solely from approved procurement records can miss publicly reachable systems for several reasons. A subsidiary may procure cloud resources independently, engineering teams may connect equipment for maintenance, or a contractor may administer infrastructure on a separate account. Internet scanning provides an outside view of what responds on public addresses, but it cannot by itself establish the full business context of each response.
The 63% AI figure measures growth in an observed internet-exposed population over a particular nine-month interval. It should not be read as a 63% increase in attacks, confirmed vulnerabilities, or successful intrusions. Similarly, the ICS count is a snapshot generated through the researchers’ discovery methods, not an enumeration of every controller used by industry worldwide.
Censys estimates that Europe accounts for 32% of the industrial control system hosts identified globally, against 38% in North America and 25% in Asia. Those figures describe the geographical distribution of observable services, not confirmed attacks on European infrastructure. Its separate AI infrastructure analysis identifies Germany as home to 31,222 publicly exposed AI-related systems. Both figures depend on the researchers’ identification methods and observation dates.
Exposure assessment also differs between an AI development interface and a legacy industrial service. AI infrastructure may grant access to models, prompts, application data or credentials if misconfigured. Industrial interfaces can connect to equipment where changes affect availability or physical processes. The presence of an exposed network service signals a possible access route, but severity depends on authentication, configuration and the purpose of that system.
The two sectors also differ in how easily defenders can establish responsibility for an exposed host. Cloud-based AI deployments may have labels, subscriptions or project accounts that can be reconciled to engineering teams. An industrial device observed through a consumer network may be administered by an installer, tenant, telecoms provider or equipment maker rather than the business whose operations depend on it. Censys' finding that about seven in ten observed industrial hosts sit on consumer or mobile networks makes ownership mapping a separate task from simply producing a list of IP addresses. The report does not prove those arrangements are unauthorised, but it demonstrates why a list of corporate address ranges is an incomplete picture of externally visible infrastructure.
Censys conducts external observation, so the findings are necessarily conditioned by what responds during the measurement period and what signatures allow researchers to identify. Services that are not publicly reachable will not appear in the same way, and a device moved behind a gateway may disappear without having been decommissioned. The measurement is therefore useful as an external-exposure indicator, not a substitute for an internal inventory.
The report arrives as organisations add AI services alongside longstanding infrastructure that is already difficult to catalogue. Its measured growth and industrial-host findings provide separate indicators of where publicly visible assets may outrun internal knowledge, while leaving actual compromise and business impact to be established through further investigation.
