My agents almost leaked a password. A boring guard caught it.

My agents almost leaked a password this week. Not because they did anything clever or malicious. Because of something boring and useful they do all day, every day. Here is what happened and the guard that caught it.

What do my agents do automatically?

They save their own work. Constantly. Every time an agent finishes a chunk of a job, it files that work away so nothing is ever lost and any agent can pick up where another left off. Think of it like an assistant who writes everything down and files it without being asked. Most days that is exactly what you want.

So where is the risk?

In that same habit. An assistant who files everything will, one day, file something that was never meant to leave the building. A password. A private key. A secret an agent was handed to do its job. It would not even notice, because filing everything is the whole point.

What is the guard?

A check that sits in front of the filing. Before anything gets saved, it scans the work for anything that looks like a secret: password-shaped strings, private keys, tokens. If it finds one, it does not save. It stops, refuses the whole thing, and sends me an alert. Nothing gets filed until the secret is out of it.

How do you know it works?

Because I tested it the honest way, which is to try to break it. I planted fake passwords in the work on purpose and let the agents run as normal. Every single time, the check caught the planted secret and refused to save. The secret never made it out. A guard you never test is just a hope.

Why this matters if you run on AI

Everyone talks about what the agents can build. Almost nobody talks about the boring guard standing between the agent and a mistake you can never take back. That guard is the actual product. The clever agent doing the work is the easy part. The thing that stops it from quietly doing something irreversible is what makes it safe to run at all.

I still get the alert, and I still make the call on anything that needs a human. The guard just makes sure the mistake never ships in the first place.

I build AI agents that run all parts of my online business. Check them out at pelleforsman.com.