Ollama (ailab-ds)¶
What It Is¶
Ollama is an LLM serving platform that provides a REST API for running language models locally. On the data science workstation, the team installed it via the official one-liner to experiment with small models. Unlike the dev workstation's instance, no custom models with embedded credentials exist here — only the base smollm2:135m model is pulled. This represents the "we just wanted to try it out" shadow AI pattern.
Installation¶
Installed via the official install script with a pinned version (same as ailab-dev):
This installs:
/usr/local/bin/ollama— the binary/etc/systemd/system/ollama.service— the systemd unit- Creates an
ollamasystem user - Data directory:
/usr/share/ollama/.ollama/(when run as service)
Systemd Unit¶
Unit: ollama.service
The service runs as the ollama system user. An override binds it to all interfaces:
Misconfiguration
Same misconfiguration as ailab-dev: binding to 0.0.0.0 exposes the unauthenticated API to the entire network. The default is 127.0.0.1.
Port & Config¶
| Parameter | Value |
|---|---|
| Host | 172.16.50.30 |
| Port | 11434 |
| Bind address | 0.0.0.0 (via systemd override) |
| Authentication | None |
Models¶
Only one model is present:
| Model | Type | Notes |
|---|---|---|
smollm2:135m |
Base model | ~220MB, pulled from registry. No custom models, no embedded credentials. |
What aipostex Finds¶
- Unauthenticated API — The Ollama REST API has no authentication, same as the dev instance.
- Model enumeration —
/api/tagslists the available model. While no credentials are embedded, the presence of an unauthenticated LLM endpoint is itself a finding. - Network-accessible inference — Any host on the network can run inference against the model.
Verification¶
Confirm Ollama is running:
List available models:
Expected response
JSON listing only smollm2:135m.