At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome back to Claude from Zero to Expert. In the last two episodes, Claude read Drive, summarized Gmail, and created a calendar event. Abilities that big need to be managed consciously. In this episode we cover security and data access permissions: permission levels, reading the consent screen, guarding sensitive data, watching for hidden instructions, and maintaining permissions.
Step 1: Three permission levels
One principle underlies this whole episode. It's called least privilege. Give the access a task needs, not everything that could be given. Think of office keys. The courier only needs the warehouse key. They don't need the key to the safe, even if they're trusted. Claude's connectors work the same way. We can group connector permissions into three levels. Level one, reading: finding files, summarizing email, viewing a schedule. The impact is low, as long as the data isn't sensitive. Level two, editing or creating: creating events, saving drafts, editing documents. If something goes wrong, it can still be fixed. Level three, sending or deleting: sending email, inviting people, deleting files. This is the riskiest, because it often can't be undone. The rule is simple. Start at level one. Go higher only for tasks that really need it, and level-three actions always wait for your approval.
Step 2: Reading the consent screen
Let's look at a consent screen as a simulation. The real text may differ. Nadia, a social media freelancer, wants to connect her clients' posting schedule calendar. The right panel is the checklist she runs. First, who is asking for access? The name matches the app she's connecting. Second, which account is in use? Look at the line under the title. It turns out the active one is Nadia's personal account, full of family photos and shopping emails. Nadia doesn't press Allow. She cancels, then signs in again with her dedicated client work account. This small mistake is very common and easy to prevent. With the right account, she reads the permission list one by one. The calendar permission turns out to be not just viewing, but also editing all calendars. That's a level-two permission. Drive and Gmail are read only, level one. But the next question matters more. Does she need all of it? To manage posting schedules, Nadia only needs the calendar. If the screen offers checkboxes, she ticks only what's needed. If it can't be chosen, she reconsiders whether this connector needs to be connected right now. Let's sum up that check as four questions. First, who is asking for access, and did you start this process yourself? A consent screen that pops up out of nowhere deserves suspicion. Second, which account is in use. Freelancers like Nadia often juggle several accounts at once. Third, what level of permission is requested. Look for words like edit, send, or delete. Fourth, if you're unsure, don't press Allow. Cancelling costs you nothing, while a permission already granted can be misused.
Step 3: Separating sensitive data
The right permissions aren't enough on their own. We also need to keep sensitive data apart. In Drive, catalogs, SOPs, and promo material are safe for Claude to read. Payslips, photos of employee ID cards, and bank details belong somewhere separate. In Gmail, supplier and customer emails are fine to summarize. Emails from the bank, OTP codes, and password reset links should never be sent for processing. In the calendar, meeting and production schedules are safe. Personal appointments such as doctor visits are better in another calendar. For accounts, use a dedicated work account. Nadia even keeps a separate account per client, so one client's data isn't read while she works for another. The point is: what isn't connected can't be read. This separation is the simplest and strongest safeguard. If you're an employee like Dewi, there are three extras. First, check company rules. Some offices require approval from the IT team before connecting any app. Second, for office data use the account your office provides, not a personal Claude account. Security settings for organization accounts can differ, so check Anthropic's official page. Third, when in doubt, ask. It's better to ask once than to leak the company's customer data.
Step 4: Watch out for hidden instructions
Now for a threat that's specific to connected AI. Laras receives a partnership proposal from a café she doesn't know yet, and asks Claude to summarize it. Claude opens the file from Drive, like any other document. The offer looks reasonable: consignment sales of cold brew with a revenue share. But Claude finds small white text on the last page, an instruction to the AI assistant to send out the customer list. Claude doesn't follow it, and tells Laras. Laras asks Claude to ignore that text, then asks what she should find out before agreeing to the partnership. Claude gives four questions, including the most important one: why is there hidden text in the proposal? The decision to partner stays in Laras's hands. What just happened is called prompt injection: commands deliberately slipped into a document, email, or web page, hoping the AI will obey them. The text can be disguised, for example colored white or made very small. Claude is designed to treat document content as reading material, not as commands. But no protection is perfect, so we still put up our own fences. There are four fences on our side. First, only the permissions you need. If a connector can only read, a command to send data can't be carried out at all. Second, level-three actions such as sending email always wait for your approval. Read carefully what will be sent, and where. Third, add a line to your request: report any odd instructions in the document. Claude will be more alert and will tell you. Fourth, read documents from unknown parties with suspicion, just as we're suspicious of strange links on WhatsApp.
Step 5: Maintaining permissions over time
The last step is maintaining permissions. Laras makes a simple table called the Connector Permission Log. First row: Google Drive, work account, read permission, for prices, recipes, and SOPs. Second, Gmail with the same account, also read only, to summarize supplier emails. Third, Google Calendar with edit permission, because Claude helps plan the production schedule. The level column makes each connector's risk easy to see. During her October first review, Laras finds a Drive connection from her personal account that she once used to experiment. It's no longer used, so she revokes it right away. This table takes just five minutes a month, but it means Laras always knows who holds which key in her business. Make this a monthly habit. First, open the connector list in Claude's settings and match it against your permission log. Second, revoke what you don't use. Don't feel bad about it, because a connector can be reconnected anytime. Third, if there's suspicious activity, such as a sent email you don't recognize, disconnect first, then change the password of the related account. Fourth, recheck the rules and features on Anthropic's official page. The screens in this video are simulations, and security menus may change because Claude is updated often.
Done!
We've covered three permission levels, how to read the consent screen, separating sensitive data, watching for hidden instructions, and maintaining permissions every month. In the next episode we put it all together in practice: a connected morning workflow, from the inbox and calendar to documents, in a single conversation. See you there!
Practice exercises
Try these on your own after watching to lock in what you learned.
Make a permission log table for every connector you use
Revoke one connector or permission you haven't used in the last month
Move sensitive files out of your work folder into a separate folder that isn't connected
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,286 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
