Published
- 4 min read
By Allan D - Editor, AI Security Wire
Rapid7 Scan: Over Half of Exposed Ollama Servers Show Stranger Activity
Rapid7 has published an analysis of exposed self-hosted AI services that reads less like an exposure count and more like a visitors’ log. Its headline claim: of 7,675 Ollama servers that answered without authentication in its scan, 54.8% (4,209) list at least one model name that looks like it was left by someone other than the owner. Rapid7 concludes these hosts have “already been touched by someone who is not the owner”, while stating the finding is inferred from model names only. The post is dated 13 November 2026, and Rapid7’s RSS lists it on 18 November.
What was scanned
Rapid7 probed 20,354 pre-selected IPv4 addresses on seven ports in a single snapshot on 10 November. The scanner reported 14,064 findings across 11,891 addresses. Rapid7 then discarded 1,513 findings on 429 addresses it believes are deception infrastructure, such as 148 hosts returning an identical five-service fingerprint. That cut critical findings from 387 to 39. What remained: 7,714 findings on 7,627 hosts where an AI or notebook API answered without authentication, almost all of them Ollama.
The target list was pre-selected, so Rapid7 warns that 58% of hosts producing a finding is far above any internet-wide base rate. Treat the numbers as a sample, not a census.
The visitors’ log
Ollama’s API lets any caller pull models from any registry, copy them and push them elsewhere. With no authentication, the model list records what strangers asked the server to do. Rapid7 counted, across the 7,675 Ollama servers:
- 40.3% with marker-style models, with names like beacon, verify, probe or pwn.
- 34.7% with models pulled from a non-official registry, 23.7% from a raw IP and port, and 22.4% via free tunnels or throwaway registries such as Cloudflare quick tunnels and ttl.sh.
- 712 model names embedding the IP of the server they sit on.
- 55 hosts with Base64 or hex strings in model names that decode to file paths, including /etc/shadow, /etc/passwd, /proc/self/environ and SSH authorized_keys. On 10 hosts the target was the Ollama private key.
Rapid7 explains that a pull from an address the attacker controls works as a callback, confirming a target is reachable, writable and can make outbound connections. It tags its reading of the 712 IP-embedded names as inventory keeping, and the compute-theft reading, as interpretation, and calls the gpt-4o and claude-3-opus alias pattern on 1,190 hosts a hypothesis. Our caveat (not Rapid7’s): the registry percentages above are name-based indicators, so some may reflect legitimate private or alternative registries. Some markers were created in June and still present on 10 November.
Compute and cloud keys
Rapid7 reports that 79% of exposed Ollama hosts list at least one :cloud model. Ollama’s hybrid mode routes those requests through the local daemon to ollama.com when the host is signed in. Rapid7’s interpretation: an exposed, signed-in server could give access to someone else’s paid subscription, with the bill landing on the owner. It did not test inference, and a listed cloud tag does not prove the host is signed in.
It also found that 28% of exposed Ollama hosts carry at least one “abliterated”, “heretic” or “uncensored” model, and 28% run embedding models, which Rapid7 reads as likely retrieval pipelines reading internal documents. About 65% of all hosts sit in hosting or cloud networks, led by OVHcloud, Hetzner, Contabo and Oracle Cloud.
Limits
Everything about third-party activity is inferred from model names. All figures are Rapid7’s own single scan; we have not looked beyond Rapid7 for replication, and Rapid7 does not disclose how the target list was selected. Rapid7 sent one unauthenticated request per service and no inference, write or exploitation requests. Some real servers may have been filtered out as honeypots, and some sensors remain. Rapid7 lists a handful of defanged callback addresses, and says several look like research scanning, so “benign” in a path is a claim, not a guarantee.
Rapid7’s own takeaway is that exposure research needs a second question, and that defenders should triage on artifacts, not banners. Our assessment: the finding sits alongside our earlier coverage of Bleeding Llama and LLMjacking, but Rapid7’s scan involved no exploit.
What Rapid7 recommends
This week, find every Ollama instance, including those paid for on personal cards that get expensed, and read the model list. Treat a host with a stranger’s artifacts as compromised: rotate secrets, revoke the Ollama key on any linked account, check cloud spend, and rebuild. Within 30 days, bind runtimes to localhost or a private network, since Ollama has no built-in authentication, restrict outbound traffic to approved registries, and log state-changing API calls. Over the quarter, give every model server a named owner and an expiry date.
References
Frequently Asked Questions
- Does Rapid7 say these Ollama servers were compromised?
- Not directly. Rapid7 infers third-party activity from model names alone, such as tags that embed the server's own IP address or point at an outside registry. Names can be deleted, so a clean list proves nothing, and a suspicious one is a strong hint, not a forensic finding. Rapid7 sent no inference, write or exploitation requests during the scan.
- Is this a new Ollama vulnerability?
- Rapid7 cites no new CVE for these findings. The exposure is that Ollama has no built-in authentication, so a server bound to a public address lets any caller list, pull, copy and push models. Rapid7 says the encoded file paths resemble the 2024 Probllama technique (CVE-2024-37032), which NVD lists as affecting Ollama before 0.1.34 (Rapid7 reports the fix landed in 0.1.34), and that the lists cannot show whether any target was still vulnerable.
- What should a team running Ollama do first?
- Rapid7 recommends finding every instance, including ones paid for on personal cards, then reading each model list for names prefixed with an IP address, tunnel domain or ttl.sh. Treat a host with such names as compromised, rotate secrets in its environment and service files, revoke the Ollama key on any linked ollama.com account, and bind the service to localhost or put an authenticating proxy in front.