How OpenAI Dots work, and what they can see in your accounts
A dot is an AI agent that runs on its own computer in OpenAI's cloud. GPT-6 Astra is the model that plans its work, and it reaches your accounts only through the apps you connect.
Before you connect a job-search inbox, know that a dot keeps notes from what it reads. At launch, the only way to clear them is to delete the dot.
The parts of a dot
| Part | What it does | What it means for you |
|---|---|---|
| GPT-6 Astra | The AI model that plans each step and decides when to ask you | Clear instructions get better results |
| Its own cloud computer | Runs a web browser and does the work away from your laptop | You can open it and watch what it is doing |
| Plugins | Connect it to the apps you choose, out of more than 4,000 | It can only reach the apps you connect |
| Memory | Keeps your ChatGPT memories plus its own notes | You cannot delete a single memory yet |
| Schedule | Runs reminders and repeat checks | It can check your inbox every morning if you ask |
| Safety checks | Look at sensitive actions before they run | Some steps always come back to you |
Where the work happens
OpenAI runs a computer for each dot in its cloud. The dot has a web browser there, so it can open sites and fill in forms while your laptop is closed. You can open that computer from the dot's profile and watch.
Your laptop is left out unless you bring it in. From the ChatGPT desktop app you can let the dot use your files and your own browser, but that access starts switched off.
Logins get extra care. For sites it supports, a dot stops and shows you a secure form, and you type your password there. OpenAI says the password goes to the dot's browser without the model ever seeing it.
What it reads, and when
A dot reads only what you connect through ChatGPT's plugins. After that, it reads your apps at these times:
- When you give it a task.
- When a reminder or repeat check you set comes due.
- When you are away. OpenAI calls this proactive research. The dot looks through your connected apps for things it could help with, using tools that can read but cannot send or change anything.
A dot can also react when something new arrives. OpenAI's example is a bug reported in Slack, which dots start looking into straight away. In a job search, the same idea could mean a reply from a recruiter that your dot flags for you.
A job-search inbox holds private details, e.g., salary offers. Connect it only if you are happy for your dot to read and remember them.
What it remembers
A dot starts with your ChatGPT memories and adds notes of its own, many of them about what it finds in your apps. According to OpenAI, those notes do not hold your login details, and they do not hold pictures or screenshots.
The memory controls are thin at launch. You cannot open one memory and delete it. Disconnecting an app stops any new reading, yet everything the dot has already taken in stays. A reset is the only full clean-up, and it deletes the dot along with its chats and scheduled tasks.
When it has to ask you
You decide how much a dot does alone. In Custom Rules, every kind of action gets one of these four settings:
- Take action without asking.
- Take action if pre-approved, which means you asked for that exact thing.
- Ask before taking action.
- Hand off to you, so you finish it yourself.
For a job search, set sending emails and messages to ask first. Then nothing reaches a recruiter until you have read it. Whatever you choose, changing a password and moving money come back to you.
OpenAI adds its own check on top, called auto-review. Before any email leaves, auto-review compares the address and the text with your instructions. A wrong address is the kind of slip it is meant to catch. OpenAI also watches for harmful behaviour while a dot works and can pause it.
Can someone trick a dot?
They can try. A web page or an email can carry hidden text written to give an AI agent orders. This is called prompt injection. OpenAI says text a dot reads does not give it permission to act, and it tested dots against this because they read new email on their own. It also says its protections lower the risk without removing it.
Fake recruiter emails are common in a job search. A dot that reads your inbox will read those too, so keep the ask-first rule on for anything it sends.
What this means for you
A dot is good at steady admin that you check, and a job search has plenty of it. For a list of tasks worth handing over, read What OpenAI Dots can and cannot do for your job search. For the interview itself, practise out loud on Just One Interview, where the report scores every answer you give.