Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Rapid7 Intelligence research banner reading New Research From Rapid7 Intelligence in green and white on a dark background.
Image: Rapid7 (fair use). Header art from the Rapid7 Intelligence report on exposed self-hosted AI servers, published Nov. 18, 2026.

Threat Intel

Rapid7 Says More Than Half of 7,675 Internet-Exposed Ollama AI Servers Already Carry Models Left by Strangers, From Callback Probes to Encoded Requests for /etc/shadow

A scan of 20,354 hosts found thousands of self-hosted AI runtimes answering without authentication, many signed in to paid cloud models and some still holding markers planted in June. Rapid7 also showed how honeypots inflate exposure counts, cutting its critical findings from 387 to 39.

Priya Shah, Threat IntelSingapore6 min read

SINGAPORE - More than half of the internet-exposed Ollama servers in a new Rapid7 scan had already been used by someone other than their owners, the company said on Wednesday, with model lists that read like a visitors' log: names that double as callback addresses, tags stamped with the server's own IP address and encoded strings that decode to /etc/shadow.

Christiaan Beek of Rapid7 wrote that a single unauthenticated, read-only request to Ollama's /api/tags endpoint returned that evidence. Of 7,675 exposed Ollama servers that survived the company's filtering, 4,209, or 54.8%, listed at least one artifact Rapid7 attributes to a third party. The report separates what it observed from what it interprets, and Rapid7 sent no inference, write or exploitation requests.

First, the honeypots

The snapshot, taken on Nov. 10, covered 20,354 preselected IPv4 addresses on seven ports, including 11434, Ollama's default. The scanner reported 14,064 findings on 11,891 addresses and labeled 387 of them critical. Rapid7 said it did not trust that number. It found 148 addresses in five countries returning an identical fingerprint of Grafana, Jupyter and Ollama on three ports, each serving one model, and another 281 that answered as two to five different products on the same port.

Rapid7 Figure 1 comparing scanner findings with findings that hold up after filtering: Ollama 8,193 flagged and 7,675 holding up, Streamlit 4,939 and 4,700, Grafana 475 and 123, Jupyter Notebook 265 and 38, MLflow 126 and 4, JupyterHub 66 and 11.
Image: Rapid7 (fair use).Figure 1 from the Rapid7 report, scanner findings before and after removing likely deception infrastructure.

Rapid7 judged those 429 addresses most likely deception infrastructure run by researchers, vendors or threat actors. Removing them cut critical findings from 387 to 39 and MLflow findings from 126 to 4. What remained was 7,714 findings on 7,627 hosts where an AI or notebook API answered without authentication, almost all of them Ollama. Only 0.3% of all findings were served over HTTPS. Because the target list was preselected, Rapid7 said the totals should be read as a sample, not a census.

Rented servers, paid accounts

About 65% of the hosts sat on hosting or cloud networks, and six providers, OVHcloud, Hetzner, Contabo, Oracle Cloud, Tencent Cloud and DigitalOcean, accounted for 42%. Germany led for exposed Ollama with 1,306 hosts, ahead of China with 1,130 and the United States with 1,097. Rapid7 read that as a map of where capable servers can be rented quickly, and warned that such machines rarely appear in an asset inventory before they are connected to company data.

Seventy-nine percent of the exposed Ollama hosts listed at least one model with a :cloud tag. Ollama's hybrid mode routes those requests through the local daemon to ollama.com when the host has been signed in, using a key stored on disk. Rapid7 said an exposed, signed-in server can give strangers use of the owner's paid subscription without any exploit, though it noted that a listed cloud tag does not prove a host is signed in and that it did not test inference. The same seven cloud models appeared together on 1,295 otherwise unrelated hosts, a pattern Rapid7 said looks more like automated pulls by third parties than individual choices.

What strangers left behind

The third-party artifacts fell into a few types. Marker models with names such as beacon, verify, probe or pwn appeared on 40.3% of servers. Models pulled from a non-official registry appeared on 34.7%, from a raw IP address and port on 23.7%, and through free tunnels or temporary registries such as Cloudflare quick tunnels, ttl.sh, bore.pub and Pinggy on 22.4%. Rapid7 said a pull from an address the attacker controls confirms that a server is reachable, writable and allowed to connect outbound. It found 712 model names that embed the IP address of the server they sit on, which it read as target inventory.

Rapid7 Figure 7 showing the share of 7,675 exposed Ollama servers with each artifact type: any third-party artifact 54.8%, marker models 40.3%, non-official registry pulls 34.7%, pulls from a raw IP and port 23.7%, pulls via a free tunnel or temporary registry 22.4%, gpt-4o and claude-3-opus aliases 15.5%, encoded file-path payloads 0.7%.
Image: Rapid7 (fair use).Figure 7 from the Rapid7 report, signs of third-party activity on exposed Ollama servers.

On 55 hosts, model names contained Base64 or hex strings that decode to file paths. The list included /etc/shadow, /etc/passwd, /proc/self/environ, Ollama's systemd unit and override file, SSH authorized_keys, AWS configuration, the Oracle Cloud metadata endpoint and, on 10 hosts, Ollama's own private key. A few pointed at /etc/ld.so.preload and new systemd service files, which are persistence locations. Rapid7 said the technique mirrors Probllama, CVE-2024-37032, which Wiz published in 2024 and Ollama fixed in version 0.1.34, and that the attempts prove intent rather than success because the lists do not show whether any target ran a vulnerable build.

Rapid7 Figure 9 showing hosts with encoded file targets by purpose: host fingerprinting 47 hosts, secrets in environment or service configuration 28, credentials and keys 15, persistence paths 11, cloud metadata endpoint 1.
Image: Rapid7 (fair use).Figure 9 from the Rapid7 report, decoded file targets by purpose.

Timestamps embedded in marker names ran from June 10 to Nov. 3, with 1,026 created in September, and some June markers were still present on Nov. 10. Rapid7 said that when a stranger's model sits on a server for five months, the larger problem is that nobody owns the server.

What to check

Rapid7 advised organizations to scan their own public ranges and every cloud and VPS account they pay for, including accounts on expensed personal cards, and to look beyond default ports, since 6% of findings sat on a nonstandard port. On each Ollama instance, it said, read the model list: anything prefixed with an IP address, a tunnel domain or ttl.sh, anything named like a beacon or probe, and anything containing the server's own IP belongs to someone else. A host with such an artifact should be treated as compromised, not misconfigured. That means rotating secrets in its environment and service files, revoking the Ollama key on the linked ollama.com account, checking cloud bills for unfamiliar usage and rebuilding the machine.

Longer term, Rapid7 recommended binding model runtimes to localhost or a private network with an authenticating proxy in front, noting that Ollama has no built-in authentication. It also urged restricting outbound traffic from model servers to approved registries, which turns a callback pull into a failed connection, and logging and alerting on state-changing API calls such as pull, create, copy, push and delete, which should be rare on a stable production server.

About the Author

Priya Shah

Priya Shah covers threat intelligence, intrusion analysis, and adversary tradecraft for SOCtember from Singapore.

Related stories

Threat Intel desk