Give an AI agent the least permission the task needs, let it read before it may write, and keep an approval where a person looks before anything is sent or changed. That is the foundation. The rest comes down to knowing what the agent can reach, being able to see what it has done and being able to shut the access quickly if something goes wrong.
Six rules for safe agents
- 1. Least permission. The agent gets access to what the task requires. An agent that compiles a weekly report needs to read the status and nothing else.
- 2. Read before write. Start with read access. Add write access only when you have seen the agent do the right thing, and then with approval.
- 3. A separate account for the agent. Give the agent its own identity in your systems, so that you can see what it has done and withdraw access without touching any colleague's account.
- 4. Approval before effect. Anything that leaves the company or changes something important passes a person first.
- 5. A trail. It should be possible to see what the agent read, what it suggested and what was approved, afterwards and in order.
- 6. A way to shut it. You should be able to shut the access quickly and know who does it.
What can go wrong
- The wrong thing is changed. The agent misreads an instruction and changes something it should not touch. Write permission with approval limits the damage.
- Access that is too wide. An agent that can read everything may bring up information that does not belong to the task. Limit what it sees.
- Text that tries to steer the agent. An email or a document can contain a line that looks like an instruction. Content in emails and documents is material to read. Instructions come only from you, and that is why the agent should not be able to send or change anything without approval.
- Information in the wrong place. Information copied into a tool nobody has approved. Your AI policy should say which tools are approved.
Questions to ask a provider
- Where is the information stored, and for how long? Ask for a written answer.
- Is there a data processing agreement? You need one when the provider handles personal data for you.
- Who at the provider can see our information? And under which conditions?
- How is what the agent does logged? Can you read the log yourselves?
- How do we shut the access? And how quickly do you learn that it is shut?
- How are incidents handled? Who contacts you, and when?
How to set permissions step by step
First write the task down in one sentence and list what the agent needs to see to do it. Then give read access to exactly that. Run it for a week and read everything the agent suggests.
When the result holds, you can add a write permission, always with approval from a person. Widen in small steps and write down every change. That makes it easy to know where you stand.
Write down every permission, who decided it and why. The list makes it easy to review and to shut what is no longer needed.
When something goes wrong
If the incident involves personal data there may be a duty to report it. Ask a lawyer or your data protection officer what applies to you.
- Shut the access. Withdraw the agent's permission straight away.
- Read the trail. See what the agent did and what was approved.
- Talk to those affected. Tell them what happened and what you are doing to fix it.
- Correct and adjust. Change the task, the permission or the approval.
- Restart in a small step. Open the access again only when you know why it went wrong.
An example: an agent that writes the weekly report
The task is to compile status from three systems into one page every Monday. The agent gets read access to the three, an account of its own and no ability to send. The result is placed in a folder where the project manager reads it.
If the agent writes something wrong it shows in the trail, and the access can be shut at once. If you later want it to send the report itself, that is a new step with an approval of its own.
Who owns security in your company
Appoint one person who owns the permissions for each agent. She knows what the agent may do, approves changes and shuts the access if something goes wrong. Without an owner, permissions remain long after the task has changed.
Go through the list of agents and permissions regularly and delete what is no longer needed.
How we work with this
With us, every step has a stop point where you look at the result before we move on. The price is fixed per step, with a cap, and is set before the step begins. We sign a data processing agreement before we touch any information.
The text here is general information and does not replace legal or technical advice for your business.
Common questions
What is least permission?
That an agent only gets access to what the task requires. If it only needs to read a calendar, it should not be able to read email.
Should an agent be allowed to write in our systems?
Yes, but start with reading. Add write permission only when the agent has shown that it does the right thing, and let a person approve what it writes.
Why should the agent have its own account?
Then you can see what it has done and withdraw the access without affecting any colleague's account.
Can an agent be tricked by text in an email?
Yes, text in emails and documents can try to steer an agent. That is why you treat content as material to read and let nothing be sent or changed without approval.
What should we ask a provider?
Where the information is stored and for how long, whether there is a data processing agreement, how the agent's actions are logged and how you shut the access.
Do we need a lawyer?
For questions on personal data and liability, yes. Ask a lawyer or your data protection officer to assess what applies in your business.