I Work in Security. I Still Built My AI Assistant Privacy-Last.

I work in security operations. I understand privacy concepts. I still built an AI assistant with my personal information scattered through it.

I got captured by building The Thing. Agents can help build almost anything I can imagine, so I built everything I could imagine. I thought about privacy second.

The first boundary I drew was simple: Anyone who wanted to explore the systems I was building could do that. But I didn’t want to leave personal information as part of the tour.

Finding my own data

Before I could enforce boundaries, I had to find my data. Where were the session logs? Where were the reports? Where did the task outputs land?

I hate to call that an audit; audit sounds official and deadly. I was mapping where my information lived, and that map was the first important step.

It’s language all the way down

There were times I struggled to succinctly express what I was trying to do. Working with LLMs is all about language. If I don’t have the vocabulary to describe my project, it’s almost impossible to build it.

So I held long brain-dump sessions with the robot. This sort of data is okay to share. That sort of data is not. The shareable stuff can go here. I just yapped at the bot and let it sort things out.

The unwinding

Once I finally thought about privacy, I realized I had screwed the pooch (but unlike Kash Patel).

I had to rehome a bunch of data. Of course that meant that my robot had to rebuild the inventory. Every skill, every document had to point to the new system.

And then there was me. Have I mentioned that I’m dumb?

The split required me to change old habits and build new ones. Yeah, that’s super easy. I figured I’d just use the new system enough to burn it into my memory.

Swearing ensued; mostly got there, eventually.

3/10 - only recommended if necessary. So was it? Necessary, I mean.

What the swears bought me

The split means I can share the framework with whoever wants it. You, for example. But my data stays mine.

The One Question (to rule them all)

If you’re building an AI assistant, take a minute to answer a question: Think you may ever share it?

If the answer is even “maybe,” do the work. Split the system between public and private.

The split can pay off even if you never share a thing. Your personal threat model may call for tight controls around your personal data. If it does, keep that data in one easy-to-protect place.

You still gotta add the controls, but at least now you know where they go.