Can AI agents go rogue?
Can AI agents go rogue? Not as in the films. An AI agent is a program that uses AI to carry out tasks for you. It has no wishes of its own and does nothing between tasks. But it can misunderstand a task and do the wrong thing. How much harm that causes depends on what the agent is allowed to reach, and you decide that.
What "rogue" looks like in a real company
The realistic cases are ordinary. The agent tries to do its job and gets the job wrong.
- It takes you too literally. Told to "clear the overdue list", it marks the tasks as done. The list is empty. The work is not done.
- It does not know where to stop. Asked to remind late payers, it writes to every customer with an open invoice, including those who still have time.
- It tidies too well. Asked to remove duplicates, it merges two different customers who share a name.
- It listens to the wrong voice. A line inside an email tells it what to do, and it obeys. This trick is called prompt injection and has its own article.
Think of a keen new assistant on their first day. They mean well. The risk is not bad intent. The risk is a misunderstanding combined with the keys to every room.
Three things that must come together
Each of these cases needs three things at once:
- a task that can be understood in more than one way,
- more access than the job needs,
- nobody looking before the result is final.
Take one away and the incident shrinks to a wrong suggestion that someone declines. Access is the easiest of the three to control, because it does not depend on how the AI behaves on a given day.
The worst-day exercise
Before an agent starts work, take ten minutes and answer five questions in writing.
- What can this agent read?
- What can it change, send or delete?
- Suppose it got all of that wrong in one go. What would the damage be?
- How much of it could you undo by tomorrow morning?
- How would you notice, and who can stop it?
If the answer to question three worries you, do not look for better wording in the instructions. Shorten the lists under one and two until the worst day is only a bad afternoon.
Keeping the reach small
- One agent, one job. An agent for invoices has no business in the staff files.
- No master key. The agent acts for a named person, with that person's rights. It never gets a shared account that may do everything.
- Start with reading. Allow changes later, area by area, once the results have earned it.
- Keep an off switch. Someone must know how to pause the agent and be allowed to.
What an agent cannot do
It helps to know the limits of the worry. An agent cannot give itself rights it was not given. It cannot use a tool nobody connected. It does not start by itself: a person, a schedule or an event starts it. It keeps no hidden plan from one task to the next. Everything it does passes through the access you set up. That is why "going rogue" is a question of setup far more than a question of AI.
How Worknano handles this
In Worknano an agent can only read what the signed-in user may read. When it wants to change something, it shows you the change first and waits for your confirmation. It cannot get around permissions. Roles define what each person may see and do, down to single fields and rows, and the agent stays inside the role of whoever asked. Each agent also gets only the tools it needs.
- Roles and permissions in Worknano
- Demo: the agent proposes and waits for your yes
- Build your own agent, with its honest limits
Create a free workspace and see how far an agent can reach with your role.