Skip to content
NewA calmer, faster ZeroHour

Exposed Ollama servers show signs of outsider compute use

Original title: Open by Default, Already Occupied: Threat Actors are Renting your Compute

Rapid7 Blogby Christiaan Beek ()13 min read
AI summary · x-ai/grok-4.7Summaries by Vani AI

After likely honeypots are removed, 7,675 Ollama findings still answer without authentication, some with callback model names.

Rapid7 says more than half of exposed Ollama servers in its scan already show use by someone other than the owner, including model names that are callback addresses, tags stamped with the host IP, and Base64 that decodes to /etc/shadow. The scanner's 14,064 findings are inflated by likely honeypots and multi-product responders; after those are removed, 7,675 Ollama findings remain, almost all unauthenticated. About 65 percent of hosts sit on hosting or cloud networks. Streamlit, Grafana, and Jupyter results are described as exposure checks, not proof of compromise.

Unconfirmed: Country rankings and some activity claims are labeled interpretation or hypothesis, not direct observation.

Key facts
  • Critical findings drop from 387 to 39 after 429 suspect addresses are removed.
  • Only 0.3 percent of all findings were served over HTTPS.
  • Six providers account for 42 percent of exposed hosts.
  • Germany leads exposed Ollama with 1,306 hosts, ahead of China and the United States.
  • MLflow findings fall from 126 to 4 after filtering.
Security assessment · AI, from the article

Impact: Unauthenticated use of exposed inference servers and planted models

ProductAffected
Ollama—
Analyze this story with Vani AIImpact, who is affected and what defenders should do, in a new chat

Indicators of compromise

auto-extracted · verify before use · export allAll
TypeIndicatorContext
domainbore.pubns of short-lived *.trycloudflare[.]com tunnels, ttl[.]sh , bore[.]pub and Pinggy endpoints used as registries. What I would cha
domaintrycloudflare.comllama:latest kimi-k2.7-code:cloud verif_sys:latest <tunnel>.trycloudflare[.]com/read/x5fa4534:b1 qwen2.5-coder:7b pwn_ctl_<this-host-ip>_
domainttl.shder:7b pwn_ctl_<this-host-ip>_ae52…:latest minimax-m3:cloud ttl[.]sh/dsv40110/dsv4probe:24h bge-m3:latest <tunnel>/L2V0Yy9zaGF
ipv4103.249.201.108glm-5.3:cloud nomic-embed-text:latest deepseek-v4-pro:cloud 103.249.201[.]108:9882/trigger/benign:latest gemma4:e4b zzvitals-beacon:lat
ipv4185.165.171.21.]107 9882 trigger/benign 43.143.113[.]68 8899 library/tens 185.165.171[.]21 8899 recon/probe Also observed: dozens of short-lived *.t
ipv424.144.109.68for hunting, not as attribution. Address Port Path pattern 24.144.109[.]68 9882 trigger/benign, trigger/plamo2-<id> 103.249.201[.]10
ipv443.143.113.68213 9882 trigger/benign 62.31.247[.]107 9882 trigger/benign 43.143.113[.]68 8899 library/tens 185.165.171[.]21 8899 recon/probe Also
ipv462.31.247.107]108 9882 trigger/benign 76.13.49[.]213 9882 trigger/benign 62.31.247[.]107 9882 trigger/benign 43.143.113[.]68 8899 library/tens 185
ipv476.13.49.213, trigger/plamo2-<id> 103.249.201[.]108 9882 trigger/benign 76.13.49[.]213 9882 trigger/benign 62.31.247[.]107 9882 trigger/benign 4

From the original article

13 min read at rapid7.com

llama3. 1:8b qwen3. 6:27b glm-5.

Rapid7 Blog

By Christiaan Beek · rapid7.com

Read the full article at Rapid7 Blog

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.