Staying safe with AI.

AI tools are the most useful — and most oversharing — software most people have ever used. The convenience has a security cost almost nobody is taught. This is the practical briefing: what's safe to share, the settings to lock down today, and the brand-new threats that appear the moment AI sits in the middle of your work.

Three short teaches, then a 10-question "smart move or risky move?" gym. About 11 minutes.

01 · Your chat is a postcard, not a vault

The habit that prevents most AI security disasters: before you paste, assume it could end up somewhere you didn't intend — a future breach, a log a support engineer reads, or data used to train the next model. Most of the time nothing happens. You only need to be wrong once.

So run a one-second test: "Would I be okay if this showed up in a leak, or in a stranger's chat?" If no, don't paste it into a consumer AI.

Never paste into a personal / consumer AI account

Passwords, API keys, access tokens, recovery codes
Card / bank numbers, SSNs, passport or government IDs
Other people's private data they didn't consent to share
Confidential work material — customer data, source code, contracts, unreleased plans — in a personal account

Convenience is not consent. A tool accepting your data doesn't make it a safe place to put it. The AI can't tell your grocery list from your company roadmap — that judgment is entirely on you.

What decides whether you're exposed

Your tier

Free/personal plans may train on your chats unless you opt out. Business/Enterprise contractually won't.

Your settings

Training, history, and feedback toggles change what's kept and seen. Defaults aren't always private.

Deleted ≠ gone

Most tools retain chats a while after deletion. Treat anything you typed as recoverable.

Match the data to the account: personal curiosity → personal login is fine; anything touching your employer or customers → the company's sanctioned, paid AI.

02 · Lock it down in five minutes

Menu names move as these products update, so treat this as "where to look" and confirm the current wording in each app:

ChatGPTSettings → Data Controls → turn off "Improve the model for everyone." Use Temporary Chat for sensitive stuff.

ClaudeSettings → Privacy → opt out of training/data-sharing. Skip thumbs up/down on sensitive chats — feedback can share them.

GeminiGoogle Account → Gemini Apps Activity → turn off to stop chats being saved & human-reviewed.

CopilotSign in with your work (M365) account for enterprise protection. Never run company data through personal Copilot.

PerplexitySettings → turn off "AI Data Retention" so your searches don't train their models.

Then lock the accounts themselves

Unique strong passwords in a password manager. Your AI account now holds a history of your thinking — treat it like email.
Two-factor auth (2FA) everywhere it's offered — this one step blocks most account takeovers.
Only official apps & extensions. Fake "ChatGPT" apps are a common way data gets stolen.

03 · The AI-era threats

Some risks are specific to AI and didn't exist a few years ago. Knowing the names is half the defense:

Prompt injection

Hidden instructions in a web page, email, or file can hijack an AI that reads them — especially one that browses or acts for you. Fix: limit an agent's access and review actions before approving.

AI phishing & deepfakes

Flawless phishing emails, a cloned voice ("your boss" approving a wire), faked video. Fix: verify unusual money/access requests on a second channel — call a known number.

Malicious AI tools

Fake "free AI" apps and extensions exist to harvest what you paste or your login. Fix: stick to well-known official tools.

Confident hallucinations

AI invents facts, citations, and code that look real. Fix: treat every output as a draft to verify — especially legal, medical, financial, security.

Oversharing through connectors

Connect AI to your email/drive/calendar and it can read & act more broadly than you think. Fix: grant least access, disconnect tools you're not using.

The one that catches people: a voice or video deepfake of a manager or family member asking for money or login help — now, with a believable reason. No real request falls apart because you took five minutes to call back on a number you trust. Slow down and verify.

04 · Smart move, or risky move?

Ten situations. For each, call it: is this a smart, safe move — or a risky one?

Question 1 of 10Score 0

—

✓

Gym complete

0 / 10

You made the call on all ten.

Your five-point sweep — do it once and it stays done:

1. Turn off model training / data sharing in every AI tool.
2. Enable 2FA on every AI account.
3. Disconnect any connected apps or integrations you're not using.
4. Keep work data in company-sanctioned tools only.
5. Delete old chats that contain sensitive information.

What you can do now

Run the postcard test, lock down settings + 2FA, spot prompt injection / phishing / deepfakes, and verify unusual requests on a second channel.

Up next

Lesson 7 · Which AI on your iPhone & Mac

Continue to Lesson 7 · Which AI on your iPhone & Mac →

Browse the tool tracks Back to AI Foundations