HostCensus

Internet-wide security measurement of publicly exposed AI inference services.

What this project is

HostCensus is an independent security research measurement project. We maintain a census of publicly reachable AI inference servers (Ollama, vLLM, LM Studio, and compatible API surfaces) in order to quantify exposure, measure remediation rates, and study the operational security of the self-hosted AI ecosystem. Aggregate findings are used for security research, including tracking how quickly exposed services are secured or decommissioned.

What our measurement traffic looks like

Our measurement node connects to publicly reachable service ports, issues a single request for public service metadata (for example GET /api/version), and disconnects. Connections are rate-limited to a maximum of ~100 new connections per second across the entire internet, and each host is visited at most once per measurement cycle.

What we never do

No exploitation, no authentication attempts, no credential testing.
No vulnerability probing beyond passive service identification.
No collection of prompts, conversations, user data, or model content.
No access to any service that requires authentication.

How to identify our traffic

Reverse DNS: scanner.hostcensus.com (forward-confirmed: the name resolves back to the measurement IP).
User-Agent: llm-expo-survey/<version> (+https://hostcensus.com/)
Measurement IP: 89.125.66.97

Opt out — permanent exclusion

To exclude your network from all current and future measurement, email vps0132@proton.me with the affected IP addresses, ranges, or ASN. Exclusion requests are honored within 24 hours and are permanent. We maintain a standing exclusion list that includes every network that has ever asked, as well as known complaint-sensitive research and honeypot infrastructure.

Contact

Abuse reports, exclusion requests, and research inquiries: vps0132@proton.me. We answer within 24 hours.