AI Assistant Privacy Basics: What Leaves Your Machine
Maker's note: 기준일 2026-10-03. We're the AZET team — we build and operate azet.io — and privacy is one of the questions we get asked most. This is generic guidance for any assistant. Where we say something about ourselves, it's dated and cautious on purpose; azet.io owns anything newer.
"Is my AI assistant private?" is a fair question with an unhelpful yes-or-no shape. The honest answer is a smaller question back: which part? Privacy with assistants isn't a single setting — it's a handful of facts about where your words go, plus a few habits that carry most of the value. Here's the plain version.
What actually leaves your machine
Three things travel when you use a hosted assistant:
- Your prompts. Everything you type is the request. There is no "just asking" tier that stays local.
- What you paste or attach. File contents, code, excerpts — the assistant can only work on what it received.
- What it reads on your behalf. If the assistant browses for you, page content and sometimes metadata ride along to wherever the model runs.
A workable rule of thumb: if you typed it and the model runs on someone else's servers, it traveled. That isn't a scandal — it's the deal. But it belongs in your head at decision time, not discovered afterward.
Local versus hosted: the honest trade
Broadly, two arrangements exist. A hosted model runs on the provider's infrastructure: strongest capability, zero setup, and your text leaves your machine. A local model runs on your own hardware: text stays home, and you pay for it in setup, hardware requirements, and generally lighter capability as of 2026-10-03.
Neither option is automatically "private." Hosted privacy depends entirely on the provider's retention and training policies; local privacy depends on your own configuration discipline. The label matters less than the behavior behind it.
| Hosted model | Local model | |
|---|---|---|
| Where your text goes | Provider's servers, over the internet | Stays on your machine |
| Capability | Generally the strongest current models | Smaller models, hardware-dependent |
| Setup | None worth mentioning | Install, hardware, updates |
| Privacy depends on | The provider's retention and training policy | Your own configuration and habits |
The questions worth asking any assistant
Before pasting anything sensitive into a new tool, it's reasonable to ask — of the vendor's documentation, not their marketing:
- Is what I send retained, and for how long?
- Is my data used to train models, and can I opt out?
- Can I delete my history, and does deletion actually propagate?
- When the assistant browses for me, what is sent where?
- Where is processing located, and who is the data processor?
You won't always get clean answers. That's informative too: the pattern of which questions a vendor answers plainly, and which they answer with adjectives, tells you a lot before you've shared anything.
Habits that carry most of the value
The high-leverage habits are unglamorous:
- Separate accounts. When an assistant connects to a service for you, prefer a dedicated login over your primary one.
- Minimal grants. Connect the one service you need, not everything the tool offers.
- Secrets stay out. An assistant doesn't need your password to help you draft an email. If a workflow seems to require pasting credentials, redesign the workflow.
- Periodic review. Once a quarter, list what you've connected and revoke what you're not using.
None of this requires understanding models. It's the same hygiene that made sense before AI assistants, applied to one more category of tool.
Where we stand, plainly
We're building azet, so the questions above are questions we ask of ourselves too. As of October 3, 2026, the assistant is in early-access waitlist at azet.io and pricing is unpublished — and we won't claim a final privacy posture in a blog post ahead of what's actually published on the product site. When there are checkable statements to make, azet.io is where they'll live, dated.
FAQ
Do AI assistants store everything I type? It varies by provider. Some retain conversations by default with an opt-out, some let you disable retention, and some don't disclose clearly. This is exactly the kind of thing to check in the documentation before pasting anything sensitive.
Are local models more private than hosted ones? Generally yes on data path — text stays on your machine — but only if your setup is sound. A misconfigured local stack can leak too. "Local" is a property of where the bytes go, not a guarantee.
Should I use an AI assistant for confidential work material? Only under your employer's rules. As of 2026-10-03, many workplaces have explicit policies about what may be pasted into which tools. If you can't find the policy, assume the cautious answer.
Following along
If you want to watch how these questions get answered in practice, the waitlist at https://azet.io is open — an email costs nothing and obligates you to nothing.