How you might know
- Files, data or settings changed that you did not request.
- Messages or posts you did not write.
- The agent says it finished tasks you do not remember asking for.
- Unusual activity right after it read an email, web page or document.
Do this now
Stop the agent
Pause or close it. If it runs in the cloud, turn off its schedule. A running agent can keep changing things.
Remove its access
Cancel and replace the keys it used. Disconnect it from email, code and databases until you understand what happened.
Read its action log
Look at what it read and what it changed. Do not rely only on its own explanation, because it can be wrong.
Assess the damage
Compare with version history and backups. Check live data and anything published or sent. If the damage is in a database, switch to the "AI agent is wrecking my database" playbook.
Recover
Restore from a backup or roll back the change. Take a copy first so you keep the evidence.
Check what it read before it acted
Look for outside text that may have steered it, such as an email, a web page or an issue. That is the usual cause.
Do not do this
- Do not trust the agent's own account of events as the only record.
- Do not leave its access in place while you investigate.
- Do not delete its logs.
Who to tell
- Whoever owns the systems it touched
- Whoever is in charge
- Customers if their data was affected
What to save before you clean up
- Tool-call and action logs
- The agent's inputs and context
- Version history of changed files and data
In the next 24 hours
- Give the agent a smaller set of permissions before turning it back on.
- Add a human approval step for risky actions.
- Write down what it could reach and what it could do.
Official help for the platforms involved
Links open each provider's own documentation, so you do not have to search for it.
What to write in the post mortem afterwards
- What access it had versus what it needed
- Which actions should have needed approval
- What guardrail you will add
Finished this one?
Add it to What I learned and earn 30 Trust. You can open your list from the badge at the bottom of the page.