Actively exploited Security news
Rapid7: thousands of Ollama servers are open without authentication, and many have already been visited
On 18 November 2026 Rapid7 published research into self-hosted model servers exposed to the internet. It found 7,627 hosts where an AI or notebook API answers without authentication – almost all of them Ollama. The bigger story is what was already on them: marker models, models pulled from raw IP addresses and throwaway tunnels, and model names that decode to paths such as /etc/shadow and /etc/ld.so.preload.
Key points
- What: Rapid7 scanned 20,354 pre-selected addresses on seven ports. After filtering out honeypots, 7,714 findings on 7,627 hosts remained where an AI or notebook API answers without authentication – the vast majority Ollama.
- Guest artifacts: 40.3% of exposed Ollama hosts carry marker models, 24% list models pulled from a raw IP and port, and 22% models pulled through free tunnels or temporary registries. On 55 hosts, encoded model names decode to /etc/shadow, /proc/self/environ and other sensitive files.
- New risk: 79% of exposed Ollama hosts list at least one :cloud model. A server signed in to ollama.com could, in Rapid7's assessment, hand strangers the owner's paid subscription.
- Only 0.3% of findings were served over HTTPS. About 65% of hosts sit on hosting and cloud networks; Germany has the most exposed Ollama hosts (1,306).
What happened
Rapid7 published research on 18 November 2026 that asks not just "is it open?" but "has someone already been here?". A scan on 10 November covered 20,354 pre-selected IPv4 hosts on ports 3000, 5000, 8000, 8080, 8501, 8888 and 11434.
After removing likely honeypots and sensors (critical findings dropped from 387 to 39), Rapid7 was left with 7,714 findings on 7,627 hosts where an AI or notebook API answered without authentication. Almost all were Ollama (7,675 confirmed findings), followed by Streamlit, Grafana, Jupyter, JupyterHub and MLflow. Rapid7 stresses this is a sample, not an internet-wide census.
Technical details
Ollama has no built-in authentication. Its documentation says it binds to 127.0.0.1:11434 by default; setting OLLAMA_HOST=0.0.0.0 exposes it to the network. Anyone who can then reach the API can pull (/api/pull), create, copy, push and delete models.
Rapid7 found these artifacts on exposed servers:
- Marker models (40.3% of Ollama hosts) – e.g.
zzvitals-beacon:latest,verif_sys:latest,ops-verify:latest,dsv4probe:24h. 712 model names embed the server's own IP address. Timestamps in marker names run from 10 June to 3 November 2026. - Models pulled from a raw IP and port (24%) – effectively a callback that tells whoever sent it the server is reachable and will fetch from outside.
- Models pulled via free tunnels and temporary registries (22%) –
*.trycloudflare.com,ttl.sh,bore.puband Pinggy. - File-read and persistence attempts (55 hosts) – Base64 or hex model names decode to
/etc/shadow,/etc/passwd,/proc/self/environ, Ollama's systemdoverride.confand/etc/ld.so.preload. This matches the old CVE-2024-37032 ("Probllama", fixed in Ollama 0.1.34 in May 2024). Rapid7 did not verify whether any target was vulnerable, so this shows intent, not success. - Cloud models (79%) – a server signed in to ollama.com could relay the owner's paid subscription to strangers, which Rapid7 calls "the self-hosted version of LLMjacking". This is Rapid7's interpretation; it did not test inference.
- Unfiltered models (28%) – "abliterated", "heretic" or "uncensored".
Who is affected
Anyone running Ollama or another model or notebook server reachable from the internet. Rapid7 notes that roughly 65% of hosts sit on hosting or cloud networks (OVHcloud, Hetzner, Contabo, Oracle Cloud, Tencent Cloud and DigitalOcean together hold 42%) and that many are "rented compute" spun up outside corporate asset management. Germany had the most exposed Ollama hosts (1,306), followed by China (1,130) and the US (1,097). Only 0.3% of findings used HTTPS.
How to tell if you were hit
List your models (curl http://127.0.0.1:11434/api/tags) and look for:
- names prefixed with an IP address, a tunnel domain or
ttl.sh, - names such as beacon, verify, probe or pwn, names containing your server's IP, and
-copy-testsuffixes, - odd copies such as
gpt-4o:latestorclaude-3-opus:latestnext totinyllama(a possible write test, per Rapid7).
Indicators from Rapid7's table (defanged):
24[.]144[.]109[.]68:9882 trigger/benign, trigger/plamo2-<id>
103[.]249[.]201[.]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[.]165[.]171[.]21:8899 recon/probe
*.trycloudflare[.]com ttl[.]sh bore[.]pub (model registries via tunnels)
zzvitals-beacon:latest verif_sys:latest ops-verify:latest dsv4probe:24h pwn_ctl_<IP>_*
Also log and alert on state-changing API calls (pull, create, copy, push, delete) – they should be rare on a stable server – and watch outbound connections from model hosts. Our IOC extractor pulls IP addresses out of your logs.
What to do
- This week: scan your public ranges and every VPS and cloud account (including ones paid on personal cards), non-standard ports included.
- Keep Ollama on
127.0.0.1or a private network. If you need remote access, put an authenticating reverse proxy or zero-trust access in front of it. - Restrict outbound traffic from model servers to approved registries and keep Ollama patched.
- Don't sign shared servers in to cloud model accounts without proxy authentication and spend alerts.
- If you find foreign models: rotate secrets in the environment and service files, revoke the Ollama key (
~/.ollama/id_ed25519) and review the linked ollama.com account, check cloud spend, and rebuild the host. - Give every model server a named owner and an expiry date; shut down the ones nobody owns.
Timeline
- 31 May 2024 – NVD publishes CVE-2024-37032 (Ollama before 0.1.34).
- 10 Jun – 3 Nov 2026 – range of timestamps in marker models found by Rapid7.
- 10 Nov 2026 – Rapid7 scan.
- 18 Nov 2026 – Rapid7 publishes the research.
Updates
- 18 Nov 2026 – first version of this article.
FAQ
Is Ollama vulnerable?
Exposure itself is a configuration problem, not an Ollama bug. According to its documentation, Ollama binds to 127.0.0.1:11434 by default and has no built-in authentication. The risk appears when OLLAMA_HOST is set to 0.0.0.0 with nothing authenticating in front. The file-read attempts Rapid7 saw target the old CVE-2024-37032, fixed in version 0.1.34 in May 2024, so they should fail on a current release.
Does a foreign model on my server mean it was compromised?
Rapid7 says to treat it as a compromise, not a misconfiguration. Someone had access to an API that can pull, create and delete models, and could have tried file reads or used an ollama.com account signed in on the box. The safe response is to rotate secrets, revoke the Ollama key, check cloud spend and rebuild the host.
How can I tell whether my Ollama is exposed?
From a machine outside your network, request /api/tags on your public IP at port 11434. If you get a model list back, it is open. Check non-standard ports too – Rapid7 says 6% of findings were on them – and every VPS and cloud account, including those paid on personal cards.
Sources
Spotted a mistake? Let us know and we will correct it openly.