Prepare before the bad day
You do not need to become a security expert. You need to know what to do when something breaks, and to give it fewer reasons to. Each item below explains why it matters and how to check it.
Your first week: five small steps
Day 1
Set a spending limit and an alert on every AI and hosting account.
Day 2
Write down where every secret key is kept, and replace any that several people share.
Day 3
Remove access you do not need: old teammates, old tools, extra admins.
Day 4
Turn on approval for risky AI actions and check what each AI helper can do.
Day 5
Restore a backup once, then book a 30-minute practice run.
Check your setup
No secret keys or passwords in your code, screenshots, chats or AI tools
- Why it matters
- Anything a secret touches can be copied. Most leaks are accidents, not clever hacks.
- How to check
- Search your project for words like key, token and password. Turn on the free leaked-secret alerts your code host offers.
Secrets are kept in a proper place
- Why it matters
- A secure place for keys limits who can read them and keeps a record.
- How to check
- Check that keys are stored in your hosting service's settings for secrets (Vercel, Netlify and others have one), not in files.
Borrowed code is locked and checked
- Why it matters
- Your app is mostly other people's code. Locking versions stops surprise changes, and a checker flags known flaws.
- How to check
- Keep the file that locks your versions with your project, and turn on a free flaw checker such as GitHub's Dependabot alerts.
Few people can change the live system
- Why it matters
- The fewer people and tools that can change it, the fewer ways it can break or be misused.
- How to check
- List who has admin access today. Remove anyone who does not need it this month.
AI helpers cannot touch secrets or delete things without a person saying yes
- Why it matters
- AI helpers act fast and can be steered by text they read. A person is the safety catch.
- How to check
- Check what each AI helper can do. Give it a practice copy of your data and require a human click for anything live.
Every risky change can be undone
- Why it matters
- After a bad change, the question is how fast you can go back.
- How to check
- Write down how to undo a change, and try it once on a practice copy.
Backups exist and you have restored one
- Why it matters
- A backup you have never restored is a hope, not a plan.
- How to check
- Restore the latest backup to a spare place and check the data is there.
Activity records are kept long enough
- Why it matters
- You cannot look into what was never recorded. Many attacks are noticed weeks later.
- How to check
- Check how many days of records you keep. Sign-ins and admin actions deserve at least 90 days; longer is better.
You know who is in charge, who decides and who talks to customers
- Why it matters
- When something goes wrong, unclear roles waste the first hour.
- How to check
- Write three names: the person who decides, the person who talks to customers, and the person who can cancel keys.
You can cancel a key, switch off a connection or pause an AI helper in minutes
- Why it matters
- How fast you can cancel is the biggest factor in how bad a leak gets.
- How to check
- Time yourself on a test key. If it takes more than five minutes, find out why.
Check how you use AI
What data can the AI helper see?
- Why it matters
- Anything it can read, it can be tricked into repeating.
- How to check
- List the folders, inboxes, databases and tickets it can reach. Remove what it does not need.
What other tools can it connect to?
- Why it matters
- Each connection is a door someone else could try.
- How to check
- Open the settings of each AI tool and review the apps and keys connected to it.
What can it do without asking?
- Why it matters
- Actions with no approval step are where accidents become emergencies.
- How to check
- Turn on approval for sending, deleting, paying and publishing.
What happens if it misunderstands the task?
- Why it matters
- AI tools are confident even when they are wrong.
- How to check
- Imagine the worst realistic misreading. Make sure you could recover from it.
Can you see what it did afterwards?
- Why it matters
- Without a record, you are trusting its own summary.
- How to check
- Confirm there is a list of its actions that you can read and keep.
Can you cut off its access quickly?
- Why it matters
- Stopping the damage depends on speed.
- How to check
- Find where its keys are kept and practise cancelling one.
Does a person review the changes it makes?
- Why it matters
- Changes it writes can contain mistakes or unsafe choices.
- How to check
- Require a person to look at changes before they go live.
Are risky changes tried on a practice copy first?
- Why it matters
- Testing on a copy is cheaper than fixing the real thing.
- How to check
- Run changes against a practice copy before they go live.
Can you undo what it does?
- Why it matters
- Every change should be reversible.
- How to check
- Keep the previous version ready to restore.
Are secrets kept out of what you type to it?
- Why it matters
- What you type to an AI tool can be recorded or shared.
- How to check
- Never paste keys into a chat. Keep them in your hosting service's secret settings instead.
Let an AI check your work
A "skill" is a set of instructions you give an AI coding assistant. Cloudflare's security-audit-skill is one for security checks. Cloudflare describes it as a skill that runs a security check in several stages and verifies each finding separately. It looks for things like tricks that fool AI tools, AI helpers misusing their tools, risky borrowed code, and accounts with too much access.
In plain steps:
- Decide what it may look at. Tell the assistant what to check and what is off limits, such as your live database.
- Let it map your app. It reads your project and lists the pieces: pages, connections to other services, and places data is stored.
- Read what it finds. Read each finding yourself. Treat them as leads, not verdicts.
- Double-check the serious ones. Confirm anything serious with a second method, or with a person who knows the system.
- Fix it, test again and write it down. Make the fix, run the check again, and note what changed.
- Add a safeguard. Add a scan, a test or a rule so the same problem is caught next time.
More tools for this, grouped by the job they do, are in the Vibe-coding security kit.
An AI check finds possible problems. A person decides what is real and what goes live. Run it on a copy, not on your live systems.