Skip to content
Security & Trust

Rapid7: model lists on exposed Ollama servers show signs of strangers

Rapid7 read the model lists of open AI servers and argues the bigger risk is that nobody owns the machine

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Rapid7: model lists on exposed Ollama servers show signs of strangers
In brief
  • Rapid7 read the model lists of exposed Ollama servers and found traces it attributes to outsiders, such as callback addresses and encoded file paths.
  • Rapid7 argues the larger problem is that nobody owns these servers. That is its interpretation, since the scan collected no ownership data.
  • Leaders can ask who approves and pays for model servers, and require an owner, a data classification and an expiry date for each.

A read-only request that works like a guest book

Ollama is software that runs AI models on a company's own hardware. Anyone who can reach an open Ollama server can ask it for its list of installed models, and no password is needed. Rapid7's researcher says that list, on thousands of machines, reads "something close to a visitors' log."

One caveat shapes everything below. Rapid7 infers outsider activity from model names alone. Names can be deleted, so a clean list does not show a clean host.

Rapid7 checked a chosen list of 20,354 IPv4 addresses once, on 10 November 2026, across seven ports. For each service it made a single request without credentials. It says it sent nothing that ran a model, wrote data or attempted an exploit.

Remove the decoys first

The scanner's raw output was 14,064 findings spread over 11,891 addresses. Rapid7 says a headline built on that figure would mislead.

The team set aside 429 addresses it believes are decoy systems, built to attract probes. A fleet of 148 returned the same fingerprint. Another 281 answered as several different products on one port. With those gone, findings the scanner rated critical fell from 387 to 39.

Rapid7 warns its filter may have removed some real servers and left some decoys in place. It also notes that 58% of the chosen hosts produced a finding, so the totals describe a sample, not the internet at large.

What remains is 7,714 findings across 7,627 hosts whose AI or notebook interface responded with no login. Nearly all were Ollama.

How a model list turns into evidence

Ollama lets a caller tell the server to download a model from any registry, duplicate or build models, and send them elsewhere. With no login in place, the model list can be read as a trail of requests Rapid7 attributes to outsiders.

Rapid7 examined 7,675 exposed Ollama servers. Some 24% list models fetched from a bare IP address and port. Another 22% list models fetched through free tunnels or disposable registries, including Cloudflare quick tunnels and ttl.sh. A fetch from an address someone else controls makes the server contact that person. That tells them the machine is reachable and can connect outward.

Some 712 model names contain the IP address of the host holding them. Rapid7 reads this as labelling, so that someone's tools can recognise the target later.

On 55 hosts, model names hide text in Base64 or hex encoding. Decoded, it points to files such as /etc/shadow, SSH authorized_keys and cloud metadata endpoints. On 10 hosts, it points to the Ollama private key.

Rapid7 says this resembles Probllama (CVE-2024-37032), a flaw Wiz published in 2024. A patched release, version 0.1.34, came out in May 2024. The lists cannot show whether any target still ran an older build. Rapid7 says the attempt shows intent, not success.

The subscription risk is unproven

Most of the exposed hosts, 79% by Rapid7's count, carry a model tagged :cloud. In Ollama's hybrid mode, requests for these models pass through the local server to ollama.com when the host is signed in to an account. Rapid7 interprets this to mean an open, signed-in server could give a stranger the use of the owner's paid plan.

Rapid7 marks this as a hypothesis. Seeing a cloud tag says nothing about whether the host is logged in, and no model was run to check. It did observe the same seven cloud models together on 1,295 otherwise unrelated hosts. It says that looks like automated pulls rather than individual choices.

Rapid7 argues the bigger risk is ownership

Rapid7 sees mostly old techniques here: an open service, path traversal, server-side request forgery and outbound callbacks. In its words, known controls "just haven't followed the workload."

It also reports that about 65% of the hosts sit in hosting or cloud networks. Its view is that a team short on compute often rents such a machine with a personal credit card, so it never reaches the company's inventory of assets. Some markers dated from June and were still present on 10 November. Rapid7 argues that if a stranger's model survives five months, the flaw is not the main problem. Nobody owns these servers.

This is Rapid7's interpretation. The scan observed no ownership data. The argument rests on how long the markers lasted and on where the hosts are rented.

The argument to take to a board is still a reasonable one. Shadow IT has a new workload and, if Rapid7's hypothesis holds, a cloud subscription attached. That second risk is still untested.

Three questions for the CFO and the CISO

First, who approved this spend? Can finance list every cloud or VPS account that hosts model workloads, including those expensed on personal cards? If the answer comes from expense reports rather than a register, the gap Rapid7 describes may already exist.

Second, what does an expiry date change in practice? Rapid7 recommends that no AI infrastructure be paid for until the server is registered with three details: who owns it, what data it may touch, and when it lapses. For a budget, the date forces a renewal, and a renewal needs a name attached. Without one, a server keeps running and billing.

Third, who may rebuild a server when its model list shows outsiders? Rapid7 advises treating such a host as compromised, rotating its secrets, revoking the linked ollama.com key and checking cloud spend. Decide who can order that before it is needed.

A server with no owner, no data classification and no expiry date is one nobody will notice when a stranger finds it.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Rapid7.

CVE-2024-37032