Is it safe to let an AI edit my site?
It's the right question to ask, and the honest answer is: on a bernard-hosted site, the safety doesn't depend on the AI being careful. It depends on what the system lets the AI do — and on the fact that nothing it does is hidden or permanent.
Why it's safe
Every change lands on a draft first. The AI doesn't type into your live site. It works on a private copy, you get a preview link, and you see exactly what visitors would see before any of it counts.
Who can press publish depends on which AI you're using. This is the part worth getting right rather than glossing:
- Bernard in your dashboard cannot publish. He edits the draft; you press Approve on the preview page, signed in to your account. That one is structural — there's no way round it.
- Your own connected AI can publish, when you ask it to. If you've connected ChatGPT, Claude or Gemini, it acts with your authority — that is what connecting it means — so it can make a draft live exactly as you would. It's instructed to do that only when you ask, and never on its own initiative.
If you want the arrangement where nothing reaches the public without your press, use Bernard in the dashboard. If you want your own AI to finish the job when you say "yes, put it live", connect it. Both are reasonable — they're just different, and you should get to choose knowingly.
The site can't run code. bernard sites are plain HTML and CSS — script-free by design, and enforced: scripts are stripped or rejected on every single write. There is no way to sneak in tracking, pop-ups or anything executable, whoever is doing the editing.
Everything is recorded, and everything is reversible. bernard keeps a full year of your site's history — every change with a plain-English reason, every publish noting what went live and whether it came from you or your assistant. Restore any version from your dashboard, or just ask the AI to put it back. This, rather than the AI being incapable, is what actually protects you.
The key is small and revocable. It covers your own sites and nothing else, and replacing it in Account → Your AI revokes the old one instantly. Treat it like a password: someone holding it could edit your drafts and publish them — replacing it is the kill switch that stops them dead, and the history is there to undo anything they did. See What is the connection key?.
Safety works in both directions, too: bernard doesn't just guard against bad changes — he actively watches for problems. Every time you publish, your site is re-checked for broken links, missing page details and anything Google would struggle to read, and every fix is re-checked after it goes live. How that works is covered in How Bernard checks your site — and what he does about it.
Safety is also about your data, not just your pages: what bernard holds, where it lives, and what the AI does and doesn't do with it. That's covered in What does bernard do with my data, and is it GDPR-compliant?.
The prompt
“Before you change anything on my site, tell me what you're allowed to do, what you're not allowed to do, whether you're able to publish, and how I'd undo anything you change.”
The [bracketed] parts are yours to fill in. First time? Connect bernard to your AI over MCP — a one-time setup in bernard → your site → Use your own AI — then paste the prompt above.
Questions people ask
- Can the AI publish changes to my website without me knowing?
- Not without you knowing — every publish is recorded. Every change lands on a draft first. If you're using Bernard in your bernard dashboard, he cannot publish at all: the live site only changes when you press Approve. If you've connected your own AI (ChatGPT, Claude or Gemini), it can make a draft live, because it acts with your authority — but only when you ask it to, and every publish is written into your version history saying what changed and when.
- Can the AI break my website or add something malicious?
- bernard sites are script-free by design — the system strips or rejects any script on every write, so neither the AI nor anyone else can add tracking, pop-ups or malicious code. And every change is recorded and reversible, so even an honest mistake costs you a few minutes rather than a disaster.
- What's the worst case if the AI gets something badly wrong?
- Before it goes live: you reject the draft and the live site was never touched. After it goes live: bernard keeps a full year of history, so any version of any page can be restored. There is no unrecoverable state.
- How much access does the AI actually have?
- It reaches your own bernard sites and nothing else — it can read them, edit them on a draft, and publish when you ask. The connection lasts until you replace it, and replacing it in Account → Your AI revokes the old one instantly. The connection key page explains exactly what the key covers.