ResearchACM IMC 2026
Ollama in the Wild: A Longitudinal Measurement of Exposed Ollama LLM Endpoints at Internet Scale
Zuyao Xu, Xiang Li, Yuqi Qiu, Lu Sun
A year-long view of exposed, self-hosted LLM infrastructure.
Research question
How do self-hosted LLM services appear, persist, and change on the public Internet? This work studies exposed Ollama deployments as a distinct Internet service, connecting endpoint availability with software versions and deployment context.
Measurement approach
The study combines daily active probing with endpoint enrichment and longitudinal analysis, covering February 2025 to February 2026.
- 152,137 cumulative IP addresses and approximately 4.8 million IP-day observations.
- GeoIP, ASN, reverse DNS, and HTTPS observations describe deployment context.
- Survival, churn, and CVE-fix adoption analyses distinguish transient exposure from persistent deployments.
What the measurements show
Exposure is persistent and concentrated in cloud and hosting infrastructure. Old versions and recurring deployment patterns remain visible throughout the observation window; the paper discusses the potential reach of provider-level mitigation.