Teaching AI Safely: A Practical Guide for Non-Technical Teams

AI Security Deep Dives, Part 10

Every piece in this series before this one has been written for a technical reader. This one isn't. If you've never heard of prompt injection and have no plans to, and you just use ChatGPT or Claude or Copilot to get through your workday a little faster, this piece is for you, and there's nothing to feel behind on. The goal here isn't to make you paranoid about AI. It's to give you a few habits that take about ten extra seconds and prevent the kind of mistake that's genuinely easy to make.

What Not to Paste

In early 2023, three separate employees at Samsung, in the span of about three weeks, pasted something they shouldn't have into ChatGPT. One was debugging equipment code and pasted proprietary source code containing real manufacturing details. Another was asked to summarize a meeting and submitted her notes, which included confidential strategy and product roadmap information. A third wanted help speeding up a slow database query and shared the actual database structure and business logic behind it. None of them were doing anything malicious. They were doing their jobs, using a tool that made the task faster, the same way most of us use these tools now.

Samsung banned external AI tools company-wide within about a month.

The categories worth being careful with aren't complicated: source code, customer information, anything with a name and a number attached to a real person, financial figures before they're public, legal documents, passwords or API keys or anything that grants access to something, and anything you've been told is under an NDA. If you wouldn't paste it into a public forum post, don't paste it into a chatbot either, the text you type usually isn't as private as it feels in the moment.

A Simple Test Before You Paste Anything

You don't need to memorize a policy document to get this right most of the time. One question covers almost every case: would I be okay with this text showing up somewhere public, or in a message to a competitor? If the honest answer is no, that's your answer about whether to paste it. If you're not sure, that uncertainty is itself the signal to pause and ask someone rather than guess.

How to Spot Shadow AI in Your Own Workflow

"Shadow AI" sounds more dramatic than what it usually is: someone using a personal ChatGPT account for a work task because it's faster than asking IT for the approved one, or installing a free browser extension that turns out to read the content of every page you're on, including internal tools and dashboards. It's rarely someone trying to cause a problem. It's almost always someone trying to get something done, reaching for whatever's available.

A quick self-check: if you're using an AI tool for work and you genuinely don't know whether it saves or trains on what you type into it, that's worth finding out before you keep using it for anything sensitive. And if you found a tool because it was free, fast, and nobody told you not to, rather than because it was the one your company actually approved, that's shadow AI, even if it feels completely ordinary.

Making the Sanctioned Path the Easy Path

If your company has an approved AI tool, know where it is before you need it, not after. If you genuinely don't know whether one exists, ask, that's a completely reasonable question and not one that makes you look behind. And if there isn't an approved option for what you're trying to do, say so out loud to whoever handles this at your company instead of quietly finding a workaround. The workaround usually isn't the problem. The silence around it is.

If You Already Pasted Something You Shouldn't Have

This happens, and it happens to careful people, not just careless ones. If it happens to you, the single best thing you can do is say something right away, to your manager, to IT, to whoever handles security at your company. The exposure from a mistake caught in the first hour is almost always smaller than the exposure from the same mistake discovered three months later because nobody wanted to bring it up. Nobody's expecting perfection here. What actually matters is catching it early.

Where This Leaves Us

None of this requires understanding how a model works, what a token is, or why prompt injection is architecturally possible, all things I've written about elsewhere in this series for the readers who want that depth. It just requires one honest question before you paste something, a little curiosity about the tools you're already using, and enough comfort raising your hand to say "I'm not sure this was okay" without worrying it'll be held against you. That's most of what good AI security actually looks like at the level most people interact with it.

I work in IT and cybersecurity, and I write about where AI capability is outpacing security and governance. This piece closes out the technical run of my AI Security Deep Dives series.