Security advice about AI tends to arrive at two volumes: “ban it all” or “it’s fine, relax”. Both are wrong. The honest position is the same as for email, cloud storage and every other tool that came before: enormously useful, worth using carefully. A handful of habits covers most of the risk.
The paste rule
Treat a chat window like a taxi: fine for conversation, not for leaving your keys on the seat. Before pasting, one-second check — would this be OK in an email to an outside supplier? Documents, code, drafts: usually yes. Passwords, API keys, customer personal data, unreleased financials: no, and there’s almost never a need — describe the shape of the thing instead of pasting the thing itself.
If you wouldn’t email it outside the company, don’t paste it into a chatbot.
Trust, but verify — especially the confident stuff
A model’s answer is a well-weighted guess, and it delivers wrong guesses with the same confident tone as right ones. For anything that matters — a contract clause, a security setting, a piece of code touching real data — the model drafts, a human signs off. AI is a brilliant first pass and a terrible final authority.
Watch what it reads, not just what you type
The newer risk: AI tools that read things for you — emails, web pages, documents. Text a model reads can steer it, and attackers have noticed; a poisoned page can whisper instructions your assistant might follow. The habit: give AI tools the least access that does the job, and keep a human in the loop wherever an AI can act — send, buy, delete — rather than merely suggest.
The starter kit
- Use work accounts for work. Business AI plans generally don’t train on your data; personal free tiers may. Know which one you’re typing into.
- Secrets stay out of chats. No credentials, no keys, no customer PII. Describe, don’t paste.
- Human sign-off on consequences. Anything that ships, spends or touches production gets human review.
- Least access for AI tools. An assistant that can read everything and act on it is an attack surface, not a convenience.
- Write the three rules down. A one-page policy people actually read beats a forty-page one they don’t.
None of this is exotic, and that’s the point: the teams that get value from AI safely aren’t the ones with the scariest rules — they’re the ones with a few boring habits applied consistently. Start there. The clever stuff can come later.