Summary
- Censys observed an average of approximately 138,000 internet-exposed hosts running industrial-control services or tooling in early 2026.
- Europe represented 31.1% of the global total, while exposure of AI and LLM infrastructure rose by more than 60% in nine months.
- The measurements indicate substantial inventory and network-boundary exposure, although public visibility does not prove that every host is vulnerable or compromised.
Europe accounts for almost a third of the industrial-control hosts visible from the public internet, while the number of exposed artificial-intelligence services has increased by more than 60% in nine months, according to new internet-measurement data.
Censys recorded an average of about 138,000 distinct hosts running industrial-control system services and tooling in early 2026, up from approximately 129,000 in 2024. Europe represented 31.1% of the global total, compared with 36.1% two years earlier.
North America accounted for roughly 38% of the observed exposure, while Asia’s share rose from 22.9% in 2024 to 27% in 2026. Europe’s percentage declined, although the region still contains tens of thousands of publicly reachable systems associated with industrial protocols, management interfaces, and supporting tools.
Censys also identified more than 294,000 distinct IP addresses exposing one of 43 AI or large-language-model tools. The corresponding figure stood at approximately 183,000 in October 2025, producing growth of more than 60% across the measurement period.
Several of the fastest-growing services have significant vulnerability histories. Exposed Langflow instances increased by 169%, while the product accumulated 18 disclosed vulnerabilities between 2024 and 2026, including four entries in the US Known Exploited Vulnerabilities catalogue. Public exposure of LiteLLM almost doubled, and an actively exploited pre-authentication SQL-injection vulnerability affects the LLM proxy.
The findings are a preview of the company’s forthcoming 2026 State of the Internet report. They are based on services observed through internet scanning and should not be read as a count of confirmed compromises.
A detected host may be intentionally accessible, protected by additional controls, inaccurately classified, or running a version unaffected by a particular vulnerability. Public visibility nevertheless establishes a reachable interface that requires an identified owner, a valid business purpose, and appropriate restrictions.
Industrial systems can remain exposed because remote-maintenance access became permanent, inherited configurations were never reviewed, or equipment was connected through consumer and mobile networks outside central IT controls. Censys said approximately 70% of exposed industrial hosts had consistently appeared on consumer or mobile networks during the preceding two and a half years.
Ownership becomes difficult to establish where an address belongs to a contractor-managed router, a remote telemetry system, a building controller, or an operational asset installed beyond the main enterprise network. External measurement can reveal systems that do not appear in configuration databases, vulnerability scanners, or procurement records.
AI services create a newer version of the same problem. Developers and business units can deploy model interfaces, workflow tools, vector databases, and LLM proxies quickly, sometimes through personal cloud accounts or default internet-facing configurations.
Those systems may contain model-provider credentials, prompts, proprietary records, uploaded documents, and connections to internal applications. An exposed LLM proxy can centralise API keys for several providers, creating financial exposure through unauthorised usage and a potential route into connected services.
European regulatory requirements place increasing weight on asset visibility, supplier oversight, vulnerability management, and the security of systems supporting essential services. NIS2 requires covered organisations to manage cyber risk across their operations and supply chains, while product-security rules place greater responsibility on vendors for secure defaults, supported lifecycles, and vulnerability handling.
Exposure management begins before patching. Organisations need to establish which services exist, who owns them, what information or functions they expose, whether public access remains necessary, and which supplier or internal team is responsible for maintaining them.
Removing an unnecessary public interface can reduce risk independently of the vulnerability state of the software behind it. Where exposure is required, access controls, network restrictions, monitoring, supported versions, and incident ownership need to be defined rather than assumed.
Europe’s reduced percentage of the global ICS total does not indicate a small estate. The combination of tens of thousands of observable industrial hosts and rapidly expanding AI infrastructure leaves a large body of systems whose resilience depends on accurate inventories and accountable deployment.
