I treat open-source AI models like unknown USB sticks

I treat OpenSource AI models like unknown USB sticks.First I scan, then I load.

A few days ago I read a small experiment on Reddit that shows why this is so important with numbers.

A reddit user (arsbrazh12) wrote a scanner for Hugging Face models.Then pointed it at 2,500 models.

86 models did not pass.

Here is what they found under the "open" surface:

• Files that only contained Git LFS pointers
• Licenses hidden inside .safetensors headers
• Models that tried to pull in extra libraries on load
• Code with patterns that also show up in malware
• Scans that failed because local tools were missing

All from normal looking repos.From places many teams see as "safe by default".

Open source gives a nice story.Transparent. Shared. Community driven.

The truth is more nuanced.

When I see teams pull a model from a hub, wire it into a pipeline and expose it to real data on day one I do not think "move fast".

I think "you just plugged an unknown USB stick into the core network".

Fast forward a few months.

You now run a "sovereign AI" platform.You talk about control, autonomy, independence.

But inside the stack you have:

• Models no one scanned
• Licenses no one read
• Dependencies no one approved

On paper you own the system.In practice you depend on whatever sits inside that model file.

Our approach at Omniance looks boring but saves sleep:

• Every external model goes through a scanner
• Every extra library goes on an allowlist, not a wish list
• Every license gets checked before first load, not after first problem

For me, an essential part of AI sovereignty starts with the question:"Do we know what we just loaded into memory?"

Have you seen similar issues? How do you handle model vetting in your workflow?