dots Custom Rules let you decide, action by action, how much freedom a dot has. Each action gets one of four modes: act without asking, act if pre-approved, ask before acting, or hand off to you. A safe starting point is "ask before acting" for anything that sends, shares, books or spends, and "act without asking" only for reading and drafting. Some limits are fixed by OpenAI and cannot be changed by a rule. This guide is based on OpenAI's published material and launch coverage as of November 2026; memory and training details come from OpenAI's dots privacy FAQ, and the four modes and fixed limits are as described by eesel and VentureBeat.
What do the four modes mean?
| Mode | What the dot does | Good for |
|---|---|---|
| Act without asking | Goes ahead | Reading, summarising, drafting, sorting |
| Act if pre-approved | Acts only within limits you set in advance | Repeated tasks to known contacts or sites |
| Ask before acting | Pauses for your approval | Sending, sharing, booking, purchases |
| Hand off to you | Stops and gives the task back | Anything you want to do yourself |
Mode names as reported from OpenAI's documentation, as of November 2026.
The difference between "act if pre-approved" and "ask before acting" matters. Pre-approval sets a standing boundary, such as replying to a known supplier. Asking means you see every instance. When in doubt, choose ask.
What can Custom Rules never override?
OpenAI has hard limits that sit above your rules:
- Money transfers always go back to you.
- Password changes always go back to you.
- Permanent deletion and software installs may need approval each time.
- Saved-card purchases need your approval.
- One approval is not ongoing permission. Approving an action once does not turn it into a rule.
- behaviour modes per action
- 4
- actions always handed back: money transfers, password changes
- 2
- auto-review step the dot cannot switch off
- 1
What is auto-review?
Before a dot acts, an auto-review step checks the planned action, for example an email's recipient and content, against your instructions, your Custom Rules and OpenAI's safety requirements. The dot cannot turn it off. It is a backstop, not a replacement for your rules: OpenAI says its defences "reduce" but "do not eliminate" prompt injection risk.
Proactive research is read-only
When a dot researches on its own initiative, that mode is read-only and enforced in code, according to OpenAI. It cannot send messages or change content while doing it.
What are sensible default rules?
- Reading and draftingAct without asking. Nothing leaves your account.
- Sending to people you knowAsk before acting at first; move to pre-approved later if reliable.
- Sending to new peopleAsk before acting, always.
- Bookings and purchasesAsk before acting; check merchant, date and amount.
- Deleting and installingHand off to you.
Write narrow rules
eesel's review reported that broad Custom Rules were rejected. That nudges you toward good practice anyway. Compare:
- Too broad: "You can send any email."
- Narrow: "You can reply to scheduling emails from people already in my contacts, with times inside my working hours."
A narrow rule limits what a misled dot can do. See least privilege for AI agents.
Avoid turning ask into always
The same review found the number of confirmations tiring. That is a real cost, and it tempts people to switch everything to "act without asking". Read why "Always allow" is risky first. A better fix is a narrow pre-approved rule for the one repetitive task that is bothering you.
What other settings affect safety?
- Local computer connection. Optional and off by default. Since 9 November 2026, a dot can connect to more than one local computer. Leave it off unless a task needs your own machine.
- Activity View. The desktop app shows what a dot is doing and lets you redirect or stop it. You can watch its cloud computer and press "Take over". See auditing your agent's activity.
- Workspace controls. Since 4 November 2026, workspace owners control which capabilities dots have. On a work account, your rules sit inside your employer's.
- Plugins. Plugin permissions are shared across ChatGPT, ChatGPT Work and Codex, so connecting an app in one may enable it in the others. Review connected apps in all three.
What about passwords and 2FA?
Saved passwords go through a secure form and aren't exposed to the model. A reviewer found a dot refused to fetch a 2FA code from an inbox it could read. Do not write Custom Rules that try to change that. Our guide to agent takeovers, passwords and 2FA explains why.
How do I know my rules are working?
Check the Activity View after the first few days with new rules. Look for three things:
- Actions that ran without asking that you expected to be asked about. That usually means a rule is broader than you intended.
- Repeated approval requests for the same harmless task. That is a candidate for a narrow pre-approved rule.
- Hand-offs you did not expect. These often come from OpenAI's fixed limits, such as password changes, and are working as designed.
Adjust one rule at a time so you can see the effect. If a dot behaves unexpectedly, stop it from the Activity View first and then review the rule, rather than changing several settings at once. Our guide to what to let an AI agent do can help decide which tasks deserve looser rules.
Bottom line
Custom Rules work best when they are boring: act freely on reading and drafting, ask on anything that leaves your account, hand off deletions, and keep rules narrow. OpenAI's fixed limits cover the worst cases, but your rules decide everything in between. For first-time setup, see our sister site's dots setup guide, and for its overall verdict, the dots review.



