A tool with hands, and responsibility that stays with us
I say this on every stage I'm on: the technology is flying, but one thing will never be automatic, and that's our personal responsibility.
From the moment Claude runs tasks, it touches your files, your email and your business. That's wonderful, as long as you know what you've given it. Anthropic says it explicitly on its safety page too: responsibility for every action Claude takes on your behalf stays with you.
So in this lesson, before we build routines: what the system does for you, and what's still on you. No scaremongering, and no lulling you to sleep.
What the system does for you
- The folder boundary. Claude reaches only the folders you've connected, and only while the desktop app is open.
- An isolated environment. Every task runs in the cloud in its own temporary environment, which is deleted at the end and never reaches your home or office network.
- Permanent deletion requires approval in every work mode.
- Computer use (Claude controlling your computer) asks for approval for each app separately.
And now the most important sentence on Anthropic's safety page: the isolation limits where Claude's code runs, not what it reads or does with the access you've given it. And one more fact to remember: files a task opens are processed on Anthropic's servers, not only on your computer.
The three approval modes
The message box has a permissions picker, and what you pick applies to the whole conversation. As of October 2026 the names vary a little between screens, but there are three modes:
| Mode | What happens | When |
|---|---|---|
| Manual (Manually approve) | Claude asks for approval before acting, and you decide on each action | A sensitive task, a new tool, or a mistake that's hard to undo |
| Automatic (Automatically approve) | An automatic safety check goes over every action and blocks the dangerous ones | Routines you've already seen run clean |
| No approvals (Skip all approvals) | Nothing checks the actions | Almost never |
And what's selected for you right now? Don't assume, look. My habit: at the start of every conversation where Claude is going to work on files or tools, I glance at the picker in the message box and choose a mode on purpose. You can switch it at any moment, and you can also stop or steer Claude in the middle of the work. And the Claude panel in Chrome has its own picker, which remembers your choice for next time.
When to choose manual, according to Anthropic: when the task touches sensitive files, accounts or websites, when you're using a tool, plugin or website for the first time, and when a mistake would be hard to undo, like sending messages or making a purchase. And my rule: start in manual, and move to automatic only after you've seen the same task run clean a few times.
Prompt injection, and the browser
What it is. Prompt injection is a malicious instruction hidden in content Claude reads during an ordinary task: a file, an email, a web page. For example, a line of white text on a white background in a document, asking Claude to send your files to someone else.
Anthropic has built several layers of protection: training the model, scanners that check outside content, and checking actions in automatic mode. Even so, it writes that the risk isn't zero. According to its explanation, an attack needs two conditions: Claude reads content from an untrusted source, and it has the ability to take an action that harms you. Remove one of them, and the attack gets much harder.
In three rules:
- A file from an unknown source goes into a separate folder with nothing else in it, and the task on it runs in manual.
- Browse only sites you know. Web content is a main channel for injections.
- A strange action is a reason to stop. Claude opens an unrelated file, or tries to send something you didn't ask for? Stop the task and report it to Anthropic with the feedback button.
The page you're signed in to
When a task needs a website, Claude uses the desktop app's built-in browser, which is separate from your own browser, or the Claude in Chrome extension. In Chrome's side panel it sees the page you're on, including pages you're signed in to: your bank, your email, your bookkeeping. Anthropic strongly recommends not using either of them for actions on sensitive information. My habit: before opening the panel, close any tabs that have nothing to do with the task. And with computer use, block sensitive apps like banking and health ahead of time.
The golden habits, and regulated professions
- A dedicated work folder. Copies, not originals.
- Sensitive information stays out. Anthropic recommends not giving access to sensitive files like financial documents. For me that also includes medical, legal and client details.
- Back up before big work. A backup isn't distrust of the tool. It's business maturity.
- Add-ons only from sources you've checked. An outside connector, skill or plugin is a supplier with a key to your office.
- A human eye before it goes out. Anything going to a client, to your accountant or out into the world.
And one more recommendation from the official page: Claude's add-ins for Excel and PowerPoint pass information between them and the task, so don't work on sensitive information in them while a task is running.
A regulated profession? Therapy, medicine, nutrition, law, accounting, financial advice: Claude doesn't know your professional body's code of ethics, and might innocently write "a treatment that will cure your anxiety". Every output that touches your field gets checked by you. It's worth anchoring this in your standing instructions too (lesson 7):
"My business is in a regulated profession. No promised results, no claims of cures, and every piece of marketing copy gets marked 'to check against the code of ethics'."
And what about the information itself? Whether your conversations are used for training depends on one toggle in your privacy settings, which we covered in lesson 2. The same toggle applies to tasks too.
Try it now
A five-minute safety check:
- Open
Trusted foldersin Settings and make sure there's no folder there with financial, medical or client information. - Go through
Customize > Connectors: every connection there should be one you recognize and need. - Open a new conversation, choose the manual mode in the picker, and send:
"Before we start working: list which folders, connectors and tools you have access to in this conversation, and exactly what you can do with each one: read, write, send or delete. Don't take any action, just describe."
- Anything on the list that surprises you, disconnect it.
- Make sure your computer has an automatic backup (for example Time Machine on a Mac), and that the training toggle from lesson 2 is set the way you chose, on purpose.
What to take from this
Anthropic's protections are good, but they don't limit what Claude does with the access you've given it. At the start of every conversation, look at the permissions picker and choose on purpose, and for a sensitive task, choose manual. And whatever you haven't connected, Claude can't break.
