SentinelOne
Endpoint and cloud security vendor whose SentinelLABS unit publishes the internet-measurement work this garden cites for the exposed-runner numbers. The January 2026 joint study with Censys, “Silent Brothers”, scanned the internet for 293 days, recorded 7.23 million observations across 130 countries and 4,032 ASNs, and found an unmanaged public layer of roughly 175,000 unique exposed Ollama hosts, a persistent core of about 23,000 doing most of the activity, and more than 48 percent of observed hosts advertising tool-calling capabilities https://www.sentinelone.com/labs/silent-brothers-ollama-hosts-form-anonymous-ai-network-beyond-platform-guardrails/. The work also describes a criminal service built on top of these endpoints, scanning for exposed Ollama, vLLM and OpenAI-compatible APIs, verifying them, and reselling access; the marketplace name Kai relays, “Operation Bizarre Bazaar,” is his rendering and was not confirmed against the primary this session. See local-ai-threat-layers, ollama.
Independent corroboration, and a useful discipline note: the Nankai University longitudinal study (arXiv 2609.07115) counted 152,137 cumulative exposed IPs across 362 days, a different number over an overlapping window, which is what internet measurement looks like when two teams count the same space with different methods https://arxiv.org/abs/2609.07115. Kai’s “175,108” is the SentinelOne-side figure; the study headline rounds to 175,000.
Sources
- “Silent Brothers”, 2026-01-29: https://www.sentinelone.com/labs/silent-brothers-ollama-hosts-form-anonymous-ai-network-beyond-platform-guardrails/
- SecurityWeek summary, 2026-01-30: https://www.securityweek.com/175000-exposed-ollama-hosts-could-enable-llm-abuse/amp
- arXiv 2609.07115: https://arxiv.org/abs/2609.07115