Rapid7 Finds Third-Party Activity Markers on Exposed Ollama Servers

Summary
A Rapid7 scan found third-party activity markers on 54.8% of exposed Ollama hosts, including callback tags and encoded file paths. The findings indicate attempted activity, not confirmed compromise, and support stronger access and egress controls.
Key points
- A snapshot scan of pre-selected addresses found 7,714 unauthenticated AI or notebook API findings across 7,627 hosts; the sample was not an internet-wide census.
- The researchers identified third-party artifacts on 4,209 exposed Ollama hosts, including marker names, models pulled from non-official registries and tags containing a host’s IP.
- On 55 hosts, encoded model names pointed to file paths including /etc/shadow, environment files, SSH keys and Ollama’s private key. The artifacts suggest attempted access, not successful reads or compromise.
- The article relates the file-path attempts to Probllama (CVE-2024-37032), but says the scan could not establish whether any target ran a vulnerable version.
- Seventy-nine percent of exposed Ollama hosts listed at least one cloud-model tag; whether those hosts were signed in or serving inference was not tested.
- The scan used read-only requests and sent no inference, write or exploitation requests. Its third-party activity indicators were inferred from model names.
- Recommended steps include requiring authentication, restricting outbound traffic, monitoring state-changing API calls, assigning server owners and rebuilding hosts if suspicious artifacts are found.
Article Details
- Attack Vectors
- Exposed Ollama APIs accepted unauthenticated requests; /api/tags exposed model lists and evidence of prior activity.
- Third parties reportedly used model pulls from callback registries to test whether exposed servers were reachable, writable, and able to make outbound connections; the article characterizes this as server-side request forgery.
- Model names on some hosts contained encoded file paths, including paths to secrets, credentials, cloud metadata, and persistence locations. The article says these indicate attempted access, but does not establish success.
- The article describes a path-traversal technique involving a model manifest’s digest field that could enable file read/write and code execution on vulnerable Ollama versions; it could not determine whether the scanned targets were vulnerable.
- Defensive Notes
- Find exposed model servers across company-owned public ranges and cloud or VPS accounts, including accounts paid for personally and expensed.
- Treat a host with third-party model artifacts as potentially compromised; rotate credentials in environment and service files, revoke linked Ollama account keys, review cloud spend, and rebuild the host.
- Bind model runtimes to localhost or a private network and put an authenticating reverse proxy or zero-trust access layer in front of remote access.
- Restrict outbound connections from model servers to approved registries to hinder callback attempts.
- Keep runtimes current; the article says Ollama fixed the described Probllama vulnerability in version 0.1.34.
- Log and alert on state-changing API calls, including pull, create, copy, push, and delete.
- Assign each AI infrastructure instance an owner, data classification, and expiry date; define where uncensored models may run and what data they may access.
Indicators of compromise
| Type | Indicator | Context |
|---|---|---|
| IPV4 | 103[.]249[.]201[.]108 | Listed as a callback registry address observed in model names; the article notes some listed registries may reflect research scanning, not confirmed attacker attribution. |
| IPV4 | 185[.]165[.]171[.]21 | Listed as a callback registry address observed in model names; the article notes some listed registries may reflect research scanning, not confirmed attacker attribution. |
| IPV4 | 24[.]144[.]109[.]68 | Listed as a callback registry address observed in model names; the article notes some listed registries may reflect research scanning, not confirmed attacker attribution. |
| IPV4 | 43[.]143[.]113[.]68 | Listed as a callback registry address observed in model names; the article notes some listed registries may reflect research scanning, not confirmed attacker attribution. |
| IPV4 | 62[.]31[.]247[.]107 | Listed as a callback registry address observed in model names; the article notes some listed registries may reflect research scanning, not confirmed attacker attribution. |
| IPV4 | 76[.]13[.]49[.]213 | Listed as a callback registry address observed in model names; the article notes some listed registries may reflect research scanning, not confirmed attacker attribution. |
MITRE ATT&CK
T1005 · Data from Local SystemEncoded model-name strings on some hosts decoded to local file paths, including system files and configuration paths; successful reads were not established.T1190 · Exploit Public-Facing ApplicationThe article describes attempts using a path-traversal payload against Ollama’s exposed API, while noting it could not confirm that scanned hosts ran vulnerable versions.T1543.002 · Systemd ServiceSome decoded paths pointed to new systemd service files, which the article identifies as persistence locations; it does not establish that services were created.T1552.001 · Credentials In FilesReported file-path targets included environment and service configuration, AWS config, SSH authorized_keys, and the Ollama private key; successful credential access was not established.T1574.006 · Dynamic Linker HijackingThe article describes code execution through corruption of /etc/ld.so.preload in the referenced path-traversal technique and reports that this path appeared among observed targets; execution was not confirmed.T1595.001 · Scanning IP BlocksThe article reconstructs an initial mass-scan of exposed model-service ports as part of the activity sequence.
CVE
Products
Grafana[Observed] 148 IP addresses return the exact same fingerprint: Grafana on 3000, Jupyter on 8888, Ollama on three ports, each serving one model, llama3.1:8b. They sit with five different hosting providers in fiveJupyterHubStreamlit, Grafana and JupyterHub results are presence checks (a health endpoint answers). They prove exposure, not compromise. Default-credential labels on Grafana were inferred by the scanner, not tested.Two moreMLflowwatching who comes knocking. Remove those 429 addresses and the critical findings drop from 387 to 39. MLflow falls from 126 to 4. Honeypots look extremely vulnerable. That's the whole job description.OllamaSample of model names returned by exposed Ollama servers. Green: left by a third party; outlined: cloud-routed or uncensored. Host-identifying parts redacted, infrastructure defanged.More than half of the exposed OllamaStreamlitStreamlit, Grafana and JupyterHub results are presence checks (a health endpoint answers). They prove exposure, not compromise. Default-credential labels on Grafana were inferred by the scanner, not tested.Two more
Countries
China[Interpretation] Germany leads for exposed Ollama (1,306 hosts), ahead of China and the United States. I don't read that as German carelessness. Hetzner and Contabo are German, OVH has a large European footprint, andGermany[Interpretation] Germany leads for exposed Ollama (1,306 hosts), ahead of China and the United States. I don't read that as German carelessness. Hetzner and Contabo are German, OVH has a large European footprint, andUnited States[Interpretation] Germany leads for exposed Ollama (1,306 hosts), ahead of China and the United States. I don't read that as German carelessness. Hetzner and Contabo are German, OVH has a large European footprint, and
