Grok Bot Can Now Use Your Phone. Here's What to Allow (and What Not To)

0:00GrokBot can now use your phone, [music] your messages, your calendar, your contacts, your location, even the lights and locks in your home. xAI started rolling it out this week on [music] iPhone and Android. Each of those apps has its own allow button. So, before you tap any of them, it's worth knowing what's actually on the other side. In this video, we'll look at what's new, where GrokBot really runs, [music] the three safety layers between it and your stuff, and the simple rule I'd use for which [music]

0:27apps to allow. Here's the update. The official GrokBot account [music] announced it can now use calendar, reminders, contacts, and more apps on your phone. That post passed 2 and 1/2 million views in its first day, [music] so a lot of people are about to see these buttons. The announcement shows six apps: [music] messages to read and send messages, location to find places near you, [music] calendar to see your day and add events, home to control lights, locks,

0:54and more, reminders, and contacts. [music] Two honest notes. When people asked, the official account said iMessage is not part of this update, and some early testers have called it [music] flaky, so expect rough edges while it rolls out. The same week, the change log added something else, [music] a vault where your bots can store logins, cards, and contact info, and use them to sign in or check out. [music]

1:18We'll come back to that. To understand what allow really means, you need to know where GrokBot lives. It's [music] not an app sitting on your phone. xAI describes it as an AI teammate that [music] works through your apps and websites on its own computer. That computer is [music] in the cloud. All of your bots share one cloud computer assigned to your account with its own browser and files, [music] and it keeps working when you close the app. According to xAI's security docs, [music] those computers are hosted by Cursor and currently run in the United States. On desktop, bots can also run commands on your own computer, and that

1:53has its [music] own setting: ask every time, which is the default, always allow, or never. Two more details [music] from those docs matter here. First, separate bots are not a security boundary. They share the same computer. [music] Second, a bot has no identity of its own. It acts as you with your access and connector [music] tokens stay on the back end, never on

2:13that computer. So, when you tap allow on your phone, you're [music] not installing something new. You're giving a bot in the cloud another set of tools that act [music] as you. That's why the six apps aren't equal. xAI hasn't published how the phone link [music] works under the hood, but you can sort the apps by what they let the

2:31bot do. Some mostly let it read, seeing your calendar, [music] finding a contact, knowing roughly where you are. If the bot gets those wrong, you get a bad answer. [music] Others let it act in the real world, sending a text in your name, adding events other people see, unlocking a door. If [music] the bot gets those wrong, someone else feels it. Keep that split in mind because it decides [music]

2:53which safety layer you should lean on. The first layer is the approval card. When an action needs [music] your okay, Grok bot shows you the exact action and its inputs with three buttons. Allow once, always allow, which can save a rule, and deny. [music] Two details are easy to miss. An approval only controls that one action. It doesn't undo anything the bot already

3:15did. [music] And if the action came from work that started without you, like a scheduled routine or another bot, the card expires after about 10 [music] minutes. Then it shows expired and the action doesn't run. [music] An expired card can still offer to always allow it in the future, which saves the rule without running the old

3:33action. [music] And one habit from the docs is worth copying. If you can't tell what an action targets or what it will change, don't approve [music] it. Ask the bot to explain it in plain language or to show you a draft first. The second layer is auto review. [music] When it's on, the bot checks its own tool calls and computer actions before they run using a model. You shape [music] it with two kinds of rules. Ask first rules always stop a matching action and wait [music] for you. Allow automatically rules let an action go through, but only if the review finds no other reason to stop. [music] And when

4:07both match, ask first wins. If a team admin turns auto review on for everyone, their locked [music] rules show up, too. And your own rules can only make things stricter. This week's update tuned it. [music] It asks less often for actions you clearly requested. It stops after you deny a card, [music] and it no longer approves actions your

4:25request said to hold. But XAI's own docs say it plainly, [music] auto review is model-based, so it should add to tight permissions and clear approval rules, not replace them. The third layer [music] is how secrets are handled. For passwords, two-factor codes, CAPTCHAs, and payment confirmations, [music] the bot hands the computer to you. You take over, finish the step, and hand it

4:47back. For some connections, it can show [music] a secure request instead. The value is masked, kept out of the transcript, and never shown to the model. The new vault [music] goes the other way. It lets bots store logins and cards, and use them on their own to sign in or check out. [music] That's real convenience. It's also exactly where you want an ask first rule, because the docs [music] list purchases, sending messages, and deleting data among the actions that

5:13deserve [music] an explicit boundary. So, here's my rule. Allow the read apps if they help you. Calendar, reminders, contacts, [music] location. For messages and home, allow them only with an ask first rule, like ask first before sending any message or before unlocking anything. [music] And keep rules narrow, like the docs' own examples. Ask first before sending any external email, or before changing a production dashboard. The docs warn [music] against broad rules because

5:40websites and tools change. Grok Bot didn't just get your phone. It got a way to act as you. [music] Now you know which buttons decide how far. If this helped, hit like so more people see it before they tap allow. And [music] subscribe because every video here takes one new piece of AI and shows

5:57you exactly how it works on the inside. [music] The announcement and XAI's docs are linked in the description.