OpenAI dots: always-on agents, what access to give

Until yesterday, an AI assistant answered when you wrote to it. With dots, introduced by OpenAI on 29 September, the assistant keeps working after you have closed the window ³.

A dot is an agent based on GPT-6 Astra: it runs in the cloud, can use a browser and connects to more than 4,000 apps ¹. It has its own computer in the cloud and carries on even when you are not using ChatGPT ².

What it can do for a business

Think of the jobs that today mean opening three programs and copying data from one to another: following an order, updating a spreadsheet, preparing a summary for Monday. These are an agent's territory.

You use it from ChatGPT, Codex, Slack and Teams ³. The first dot is included in the Pro and Business Premium plans in eligible markets ², and for Enterprise it is in trial if the administrator enables it ¹. The sources do not say which countries are eligible: if you work in Italy, check in your account before you plan anything.

Where the risk is

The risk with an agent is not that it gets a sentence wrong. It is that it does something: sends an email, changes an order, deletes a file. OWASP, which catalogues software vulnerabilities, calls this "excessive agency": too many functions, too many permissions, too much autonomy ⁴.

The timing is instructive. In the same days OpenAI decided not to launch another model because it tended not to disclose some of its actions to the people supervising it ⁵. If the makers themselves struggle to keep the most autonomous models on a leash, the leash on your side should be kept short.

The controls to set before you connect it

According to the sources, dots offers per-app permissions, custom rules and approvals for some actions, and in the background it cannot send messages or edit content ¹. For businesses there is also integration with Microsoft's security controls ³.

Use them like this, following OWASP's recommendations ⁴:

• Connect only the apps the task needs. An agent that summarises orders does not need your email.
• Give the lowest permission possible. Read before write; write to one folder, not to the whole drive.
• Have a person approve the actions that matter: payments, emails to customers, deletions.
• Decide who can switch it on. If you have a company plan, the choice belongs to the administrator, not to the individual employee.
• Check the action log for the first few weeks, even if everything seems to be going well.

What to do now

If nobody in your business uses ChatGPT on a Pro or Business Premium plan, you can wait. If someone does, today is a good day to write a couple of lines on which apps an agent may touch and which it may not. It is easier to say it beforehand than to explain it afterwards.