Your AI agent went wrong.
Work from the top and tick as you go. The clock starts on your first tick. Nothing you type here leaves this page.
Past the hour. Keep going in the same order.
Do these first
Before you read anything else.
Three questions
Your answers cut the list down to what applies. Not sure? Leave them blank and do everything.
Where does it run
What could it reach
What happened
T+00 to T+05NOW
Stop it
T+05 to T+20NOW
Cut its access
Assume every key the agent held is burned. Revoke now, replace later.
T+20 to T+40NOW
Find out what it did
Read, don't fix. You are making a list.
T+40 to T+60NOW
Tell who needs to know
Short and plain. You are warning people, not confessing.
The note
It writes itself as you tick. Paste it to your bank, your IT team, or whoever helps you.
Clear the clock, every tick and the note?
After the hour
Before you turn it back on
Six habits from running an agent with real access every day.
- Its own accountsGive the agent its own email, its own keys and its own machine or user. Never yours.
- One key per jobShort-lived where you can. When something leaks, you revoke one thing, not your life.
- You approve what can't be undoneMoney, sends, deletes and anything public wait for a person.
- Read add-ons before you install themA skill or plugin runs with everything the agent can reach.
- Never on the open internetReach it over a VPN or private tunnel with a password. No bare ports.
- A log you read, a stop you've testedLook at what it did each day, and practice shutting it off once before you need to.