Strangers already squat on half of exposed Ollama servers

Published

Ask an exposed Ollama server which AI models it has installed and, more often than not, the answer includes models its owner never put there. Some are named after the server's own IP address. Some are really callback addresses. A few are Base64 strings that decode to /etc/shadow.

That is the finding of a scan by Rapid7's Christiaan Beek, published in Rapid7's report on self-hosted AI exposure. Of 7,675 Ollama servers that answered without authentication, 4,209 (54.8%) showed at least one sign of third-party activity. As Beek puts it, these lists describe activity, not just risk: an unauthenticated read of the model list returns something close to a visitors' log.

The number Rapid7 refused to headline

The raw scanner output was far scarier: 14,064 findings across 11,891 IP addresses, 387 of them rated critical. Beek threw much of it out. 148 addresses in five countries returned an identical fingerprint (Grafana, Jupyter and Ollama, each serving the single model llama3.1:8b), and another 281 answered as several different products on one port. Rapid7 reads those as likely honeypots and sensors. Removing them cut the critical findings from 387 to 39. What survived was still 7,714 findings on 7,627 hosts where an AI or notebook API answered with no authentication, almost all of them Ollama, and only 0.3% served over HTTPS.

The scan was a single snapshot on 10 November 2026 against 20,354 pre-selected addresses, so Rapid7 treats the totals as a sample, not a census. Roughly 65% of hosts sat in hosting or cloud networks, with OVHcloud, Hetzner, Contabo, Oracle Cloud, Tencent Cloud and DigitalOcean accounting for 42%. Germany led for exposed Ollama with 1,306 hosts, which Beek attributes to where cheap servers are rented, not to German carelessness.

A free route into someone else's AI bill

Ollama's API lets any caller pull models from any registry, copy them and push them elsewhere. Rapid7 found 24% of hosts listing models pulled from a raw IP and port, and 22% pulled through free tunnels or throwaway registries such as Cloudflare quick tunnels and ttl[.]sh. A pull from an address the attacker controls confirms the server is reachable, writable and able to call out. 712 model names embedded the server's own IP, which Beek interprets as somebody tagging targets for later.

On 55 hosts, encoded model names pointed at /etc/shadow, /proc/self/environ, SSH keys, AWS config, Ollama's systemd override file and, on 10 hosts, Ollama's own private key. Rapid7 links this to the Probllama path-traversal technique (CVE-2024-37032), fixed in Ollama 0.1.34 in May 2024, and is careful to say the attempts prove intent, not success.

There is a money angle too. 79% of exposed hosts list at least one cloud-routed model. If a server has been signed in to an ollama.com account, Rapid7 says a stranger may be able to ride the owner's paid subscription, a self-hosted form of LLMjacking that needs no exploit. Rapid7 did not test inference, so this remains unconfirmed per host. It is a different abuse from the cryptominers hijacking AI GPU servers, but it targets the same kind of forgotten box.

Read the model list, then lock it to localhost

Callback registries Rapid7 saw most often include 24[.]144[.]109[.]68:9882, 103[.]249[.]201[.]108:9882, 43[.]143[.]113[.]68:8899 and 185[.]165[.]171[.]21:8899. Rapid7 notes several look like research scanning and offers them as hunting context, not attribution.

Some markers were planted in June and were still there on 10 November. The attackers here did not need AI at all, just an open port and an owner who stopped looking.

This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.

Read the full analysis on IntelFusions