RunCabin Blog · How it works

Is it safe to let an AI agent update your website?

October 14, 2026 · 9 min read

Letting an AI agent update your website can be safe, if the host keeps a person in charge of three things: who gets in, what costs money, and what can be undone. On RunCabin, an AI agent reaches only the sites on your own account, every publish is kept so one request rolls a bad one back, and anything paid is a checkout link you open, review and pay yourself. Here is what an agent can do on its own, what needs you, what a rollback can't undo, and what it can't touch at all.

If you build or run agents, the short version with setup steps and prices is on our page about website hosting for AI agents. If you just want to keep your own site current by asking, start with how to connect and manage your website with AI.

How does an AI agent publish a website?

There are five steps, and only the first one is yours:

1. You approve the connection once, in your own browser.
2. RunCabin tells the agent to read its build guide, the rules for what a page needs and what it must never make up.
3. It writes the files and publishes them in one request.
4. The reply carries the live https address and a list of checks: anything the page is still missing.
5. It fixes those and publishes again to the same address.

A custom domain comes later: the build guide tells agents to add one only when you ask, and it takes a checkout you pay for yourself. The technical steps are on our agent hosting page.

What can an AI agent do to your website on its own?

On RunCabin, an agent works through a fixed set of tools, each labeled as either reading your site or changing it. What your AI app does with those labels is up to that app, so here is what each request can do without you:

Anything that costs money or grants access needs you, and a bad publish can be rolled back. Four other changes in the table can't be; they are listed under what a rollback can't undo.

Does an AI agent need your password to update your website?

No. On RunCabin, giving an AI agent access to your website doesn't mean giving it your password: an agent never needs or sees your RunCabin password.

Chat apps and coding tools such as Claude, ChatGPT, Claude Code and Cursor send you to RunCabin's own sign-in page in your browser. The approval screen there shows the app's name and where it sends you back. If RunCabin doesn't recognize the app, the screen warns you before you approve anything.

Scripts and command-line tools that call RunCabin's REST API can use a key you create at runcabin.com/keys, or device login, which needs nothing pasted. With device login, the tool shows you a short code; you open runcabin.com/keys, sign in, check that the code matches and approve it, and the tool picks up its own key. A request expires after 10 minutes and works once. When a tool gets a key this way, we email you the name it gave, its key prefix and the time. A key is shown once and stored on our side only as a hash. Setup steps: sign in from a terminal.

The approval page at runcabin.com/keys carries a warning worth remembering: "RunCabin never emails or texts you a link like this." If someone sends you a code to approve, deny it. To remove an agent's access, revoke its key or connected app at runcabin.com/keys. A revoked app is signed out everywhere immediately.

Can an AI agent buy a domain or spend your money?

An agent can find a domain and hand you a checkout link, but RunCabin never charges you through the agent. Checking whether a name is available, and its price, is free. RunCabin's tools stop at the link: they never take a payment themselves.

• A new domain. The link opens RunCabin's order page. It shows the domain's price and hosting at $4.99 a month or $47.99 a year, whichever you pick; you sign in and pay. After payment, registration, DNS and HTTPS are set up automatically.
• A domain you already own. The link starts the hosting subscription. It expires after 24 hours, and the agent can make a new one. Once you subscribe, you add the records it gives you at your domain provider.

Nothing is charged until a checkout is finished and paid. A domain bought through RunCabin is yours; if you cancel, we release it to you with its transfer code, as our terms describe.

Can you undo a change an AI agent made to your website?

Yes, if the change was a publish. A bad publish is the easy case: a wrong price, a section that vanished, a layout that broke on phones. Every publish is a new version and earlier versions are kept, so rolling a bad one back is one request, such as "put it back the way it was yesterday." The earlier version serves again at the same address.

An agent can also stage a version at a preview address and make it live only when it looks right. If a first publish fails, the reply still names the site it created, so the agent can try again on that site instead of making a second one. And the build guide tells agents to publish only when you want the site online.

What a rollback can't undo

A rollback changes which published version is live. It doesn't reverse these:

• Deleting a site. It is permanent, and the site's form messages are deleted with it.
• Replacing an uploaded image or font. Uploads aren't versioned, so a rollback doesn't bring the old file back.
• Swapping your custom domain. The old domain stops serving the site, and rolling back the pages doesn't reattach it.
• Changing where email forwards. Mail goes to the new address from then on.

Those are the requests to read before you say yes.

How do you stop an agent from making things up on your site?

The bigger risk with AI on a business page isn't a typo. It's a confident claim you never made: a review, a license, a price.

Before it builds, the agent is told to read RunCabin's build guide: only facts you gave, never invented reviews or ratings, forms wired to a real address, alt text on images. After each publish, the reply lists what the page is still missing, while the site is already live: a placeholder left in, a form that posts nowhere, an image without alt text, a missing page title. The guide tells the agent to fix the list and publish again. The checks catch missing pieces, not false claims, so read the live page before you share the address. Of these checks, one blocks a publish outright: a file that contains a recognizable secret key, such as a Stripe, AWS or OpenAI key or a private key, is refused before anything goes online.

More on what the guide tells agents and the standard your AI builds to.

What can't an AI agent touch on RunCabin?

• The server. No server access and no database: RunCabin hosts these sites as static HTML, CSS and JavaScript files only.
• Anyone else's sites. A key or sign-in reaches only its owner's sites, domains and forms.
• Most of your DNS. On domains bought through RunCabin, the agent can turn on email forwarding and add new TXT records; it can't edit or remove the records that keep your site serving. On a domain you connect, you add every record yourself.
• Your forms as a back door. Through RunCabin's MCP server, messages visitors type into your forms reach the agent labeled as untrusted visitor text, so it can tell them apart from your instructions. A label helps, but it isn't a guarantee.
• Private addresses. Images, fonts and other uploaded files are fetched only from public https links, never from internal network addresses.

Five questions to ask before you connect an agent to any website

These apply to any host, not only ours. If a host can't answer one plainly, think twice.

1. Can it spend money, or only hand you a checkout? On RunCabin, every paid step is a checkout link you pay yourself.
2. Where do you approve it, and where do you take access back? You approve in your own browser and revoke any key or app at runcabin.com/keys.
3. Are earlier versions kept, and what can't be undone? Every publish is kept, and a bad one rolls back in one step; deletes, replaced uploads, domain swaps and forwarding changes stay done.
4. What can it never touch? Other people's sites, any DNS beyond adding TXT records and email forwarding, and server code.
5. Does it check its own work before and after it publishes? RunCabin tells the agent to read the build guide before it builds, and every publish comes back with a list of fixes.

What do you still do yourself?

An agent takes the busywork. These stay with you:

• Approve the connection.
• Give the facts the page should state.
• Decide when the site goes live.
• Pay at checkout.
• Read any request to swap your domain, change where email forwards or delete a site before you say yes.
• For a domain you already own, add the records at your provider.
• Revoke keys you no longer use at runcabin.com/keys.
• Cancel hosting at runcabin.com/cancel when you're done.

Where does the site run?

The sites run in Hetzner's Falkenstein data center park in Germany, which draws 100% renewable electricity from certified hydropower. There is no green tier and no extra cost. We make no carbon claim beyond that. Our about page shows how we checked the data center's power, and our post explains why we moved to a green data center.

Green data center· 100% hydropower

---

Related reading: website hosting for AI agents · how to connect and manage your website with AI · how to publish a website from Claude · how to put your website online from Muse · setup for each app