Lesson 7 of 9
Assume it already happened
Plan for the night an agent goes wrong: a wrong send, a delete, a leak. The questions are how you would know by morning, what it could reach, and how you get back to yesterday. Many of the costly failures on record had no attacker at all.
We tell our kids not to take photos of themselves with their phones that they wouldn't want printed and hanged in their school's public board.
The same applies to security breaches. Start by assuming that you were breached. And now figure out backwards what's needed.
I take an image backup every night. Someone else added that the best defense is a system that tells you something unusual is happening. Someone else said git saved them many times. And one warning worth repeating: make sure no exported chat history goes to git. Those are other people's messages.
Many costly failures had no attacker
Many of the costly failures on record with agents had no attacker at all. One widely reported case: an agent mass-deleting email while its owner typed stop. Imagine the quieter versions too: a folder deleted because it looked like clutter, a message sent to the wrong list. Plan for the night your own agent goes wrong, and you have planned for most attacks as well.
Three questions for the morning after
- How would I know by morning? One line per outbound action your agents take, written somewhere they cannot edit, and five minutes a week reading it for anything you did not expect
- What could it reach? The access it has now, which is usually more than the job needs (lesson 3)
- How do I get back to yesterday? A backup you have restored from at least once. A backup you never restored is a hope
To be honest, for my own use, my agents have access to almost everything. They can't move money and that's likely the only thing they really don't have access to. My main worry is getting back to yesterday.
For a business, the answers to the above will be different.
Other people's data
Your workspace probably holds other people's words: exported chats, client emails, meeting notes. They do not belong in a git repository with a remote copy, in a synced folder, or in front of an agent that can post. Know which folders hold them and which agents can read them.
The prompt
Answer from what exists today. This prompt is about the system you have, before anything new gets built.
If one of my agents went wrong tonight, sent the wrong thing, deleted a folder or posted my private notes, answer from what exists today: 1. how would I know by morning? 2. what is the worst it could reach with the access it has now? 3. how would I get back to yesterday's state? Then tell me which of my folders hold other people's data and whether any of them is in git.
Reading the answer
A real answer
Three honest answers, including "you would not know" if that is the truth, and a list of folders.
A non-answer
"We can build monitoring for that." That answers a question about the future. Ask about tonight.