How to keep Claude safe when it can take actions

When Claude can edit files, send emails or run code, you need guardrails. These are the controls that stop the tool doing something harmful by mistake.

This guide shows you how to set them up.

Why this matters

If Claude only writes draft text, the worst it can do is write badly. But if it can move files, post updates or change data, a mistake could cost you time or money.

Guardrails let you give Claude useful powers while keeping control.

How it works

Claude has two main ways to stay safe when it acts.

The first is approval mode. You choose which actions Claude must ask you about first. For tasks that touch sensitive files, accounts or tools where mistakes are hard to undo, you approve each action yourself. For tasks you trust, you can switch to automatic approval. Even then, Claude reviews each action for safety before it runs. It always asks for explicit permission before it deletes a file for good.

The second is access limits. You control what Claude can reach. Instead of giving it access to your whole computer, you give it a dedicated working folder. Instead of letting it visit any website, you limit it to sites you trust. Instead of letting it use any tool, you stick to verified extensions.

In Claude Code (a version of Claude built for writing and running code), you can also set up permission allowlists. These let you pre-approve only the actions you trust. You can also use hooks, which are small scripts that run automatically at set points, such as before Claude makes a change.

How to do it

1. Choose a working folder for the task. Move only the files Claude needs into it. Do not give Claude access to folders that hold financial documents, passwords or personal records.

1. Decide which actions need approval. Start with approval mode on for any task that could send messages, change shared files or touch customer data.

1. Test the task once with approval on. Watch what Claude tries to do. If every action looks safe, you can switch to automatic mode.

1. For internet access, name the sites Claude may visit. Do not give general web access unless the task needs it.

1. If you set up a scheduled task, review the past runs once a week. Check that Claude is still doing what you expect.

An example prompt for a task with guardrails built in:

> Draft replies to support emails in the folder called Inbox. Save each draft as a text file in the folder called Drafts. Do not send anything. Do not open any links in the emails.

Common mistakes

Giving Claude access to a whole drive when it only needs one folder. This means a mistake could touch files that have nothing to do with the task.

Switching to automatic approval too soon. Run the task a few times first so you see what Claude does.

Ignoring a safety feature called prompt injection. This means hidden instructions placed in a web page, email or document that Claude reads as part of a normal task. It needs two things at once to work: Claude can read content from outside your trusted area, and Claude can take actions that could cause harm. The fix is to limit both. Only let Claude read from sources you trust, and keep approval on for anything risky.

Forgetting to check scheduled tasks. If a task runs every day, a small mistake can multiply.

One thing to try this week

Pick one task you would like Claude to help with. Set up a working folder with only the files it needs. Add a line to your prompt that tells Claude what it must not do, such as do not send anything or do not open links. Run the task once with approval on and check every action before you allow it.

The Claude Help Center has full guidance on using Claude safely.

Automating more tasks can save your business time. Take the free What could your business automate? check.