On Instagram and in short videos, I keep seeing tutorials that say you can build your own AI just by pasting a link, or that adding one thing will let your AI do far more. Watching them, I want to try it tonight. I recently saved two of those videos myself.
But short videos usually show you how to set things up without explaining what is on the screen that asks for your permission halfway through. Today I have put together three things to check before you try, based on OpenAI’s official documentation.
What you will be able to do after reading
-
Read a permission request and understand what you are allowing
-
Know how buying extra usage works when you hit your limit, so you decide your own spending
-
Check for yourself how the information you enter is handled
Hi, I’m Kawano from eager. Alongside my day job, I build websites for small shops with AI (three for real shops and two demos for fictional ones; I have not yet received or delivered a paid project). I can’t write code. That is exactly why I want to decide in advance what I hand to AI and what I check myself.
To be clear, this article is not against those tutorial videos. I do recommend trying things. I just think you can try with more peace of mind if you know what to read before you click.
1. Permissions: Read What You Are Allowing Before You Click
What the AI can do without asking, by default
A common setup in recent tutorials is to choose a work folder in Codex (OpenAI’s AI that writes and runs code) and let it work inside that folder.
According to OpenAI’s documentation, Codex with default settings can read and write files and run commands inside the folder you chose, without asking. On the other hand, it asks for approval in cases like these:
-
When it wants to change files outside the work folder
-
When it needs to connect to the internet (internet access is off by default)
-
When it wants to do something that can’t be undone (an action that could break things)
In other words, when you are asked for permission, the AI is either trying to step outside its set boundaries or trying to do something that can’t be undone. Instead of clicking through the way the video does, take a moment to read which folder it is touching and what it wants to do.
Make a new, empty folder to work in
Since the AI can read and write inside the folder without asking, I think it is best to create a new folder with nothing in it when you try this. Don’t pick a folder that holds your work documents or photos.
Don’t switch to “Full access”
There is a setting that removes approvals and the boundaries too (Full access). It looks convenient because you no longer get asked every time, but OpenAI says in its documentation that it does not recommend it. When trying things out, I think the default settings are enough.
When you connect apps to ChatGPT
ChatGPT has “apps” that connect it to other services (they used to be called “connectors”; the list has now moved into the “plugins” list). According to the documentation, an app can see what the connected account can see. Some apps can even create or update things in the connected service.
When you stop using an app, you can disconnect it. After trying something, it is reassuring to check that nothing is still connected.
2. Money: Know How “Buying More” Works When You Hit Your Limit
According to the documentation, Codex is available on every plan, including the free plan. But how much you can use depends on your plan.
On the paid Plus and Pro plans, when you reach your usage limit, you can buy “credits” and keep going without upgrading your plan. That is handy, but when you are absorbed in your work, you might click the buy button without thinking it through.
When you try it, I recommend deciding these two things in advance:
-
If I hit my limit, I stop for the day
-
Before buying more, I check the price on the pricing page
Limits and prices can change, so I haven’t put any numbers in this article. Please check the latest amounts on OpenAI’s pricing page.
3. Data: Check the Setting That Keeps Your Information Out of Training
ChatGPT has a setting that keeps your conversations from being used to train its AI. According to the documentation, if you turn off “Improve the model for everyone” under Data Controls in Settings, your conversations stay in your history but are not used for training. I haven’t checked the actual screen myself; the setting name is taken from the documentation.
One thing to note: Codex has a separate setting too. On a personal plan, your Codex work is covered by the setting above. But the Codex settings screen has a separate option for whether your whole work environment can be used for training, and the documentation says turning off the ChatGPT setting does not change it. Please check both.
Better yet, don’t enter customers’ names or contact details at all while you are trying things out.
Just One Thing About Links From Strangers
Some tutorials say all you need to do is paste a link and ask the AI to make it usable. The link usually points to a place where someone has stored a program they made (such as GitHub).
OpenAI warns in its documentation that AI can end up following instructions hidden in text it reads (this is called prompt injection). In June 2026, it was reported that a proof of concept had been published showing that even a code repository that looks harmless can hijack an AI agent (the report said no real-world damage had been confirmed).
In Japan, IPA (the Information-technology Promotion Agency, a government body) published its “10 Major Information Security Threats 2026,” and “cyber risks around the use of AI” (my translation of the Japanese names) appeared on the list for organizations for the first time, at number 3 (it is not on the list for individuals).
I don’t want to scare you. My realistic line is this: if you want to try a link from someone you don’t know, don’t do it on the computer or in the folders you use for work.
Checklist Before You Try
You can copy this and paste it into your notes app.
-
I created a new, empty work folder
-
When asked for permission, I read which folder it touches and what it will do
-
I don’t switch to Full access
-
After trying, I disconnected any apps I connected
-
If I hit my limit, I stop for the day. I check the price before buying more
-
I checked the “don’t use for training” settings in both ChatGPT and Codex
-
I don’t enter customers’ names or contact details
-
I don’t try links from strangers on my work computer
Note: This is based on OpenAI’s official documentation (in English) and IPA’s page as of November 1, 2026. Setting names and defaults may change.
Note: I did not operate and test each of the settings above for this article. It is a summary of the official documentation.
