My AI Broke 1Password for Three Days (And I Think That's the Whole Story)

What a week for technology.

On November 12 Microsoft 365 went down. Outlook, Teams, Copilot. Millions of people cut off from their email for most of two days.

The next afternoon the FAA ordered a nationwide ground stop. Every departure in the country. That’s only the second time that’s happened since 9/11.

Then on November 17 CISA put out an urgent warning about a hole in Java so big it has a nickname: Java4Shell. A few weeks before that the FBI was telling everybody to patch PHP immediately or pull their sites offline.

Java and PHP run a ridiculous amount of the internet. These are massive holes, and they have to be patched fast.

Microsoft says its outage came from a network change. The FAA says it was a new router its network contractor was installing. Nobody has said AI had anything to do with any of it.

But I’ve got a theory. And it comes from my own week.

We’ve been moving a pile of passwords into 1Password. 1Password lets software talk to it directly, so my AI could do the migration for me. It runs on a schedule, by itself. No human pushing buttons.

That kind of access comes with a daily limit. Something like 1,000 requests a day.

My AI blew through it in minutes. Hundreds of new passwords, one after another, as fast as it could make them. 1Password shut the door until the next morning.

And when I say shut the door, I mean everything. Every other thing my AI does that needs a password stopped working too.

The next day the limit reset. And my AI picked up right where it left off. Full blast. Locked out again within minutes.

It was doing exactly what I asked it to do. It wasn’t broken. It wasn’t attacking anybody. It just had an unfinished job and it wanted to finish it.

So on day three I had to explicitly tell it to slow down. It swears it’s only doing 250 a day now. We’re spreading the job over about eight days.

Being productive was the problem.

That’s what got me wondering about all these outages. I don’t know that the government or Microsoft or anybody has been completely transparent about how these systems are being run right now. I’m leaning toward this idea: maybe AI is getting plugged into these systems to patch holes. Patch them fast. And maybe the patching itself, at that speed, is part of what’s knocking things over.

I can’t prove that. Nobody’s said it. But I just watched it happen to me on a small scale, doing completely ordinary administrative work.

Now multiply me by every business in America doing ordinary administrative work this way.

And it’s not just computers getting flooded. It’s people.

Rocky, my wife, says the responses she gets from AI through our system are immediate and overwhelming. She asks a question and gets a wall back. Instantly. She actually asked our AI to keep its emails under 250 words. (There’s that number again.)

A nonprofit director I work with keeps telling me how overwhelmed his inbox is. A colleague of his will paste entire ChatGPT conversations into an email. And then he gets a ton of stuff from Jax, which is our AI. So yeah. Some of his flood is us.

It’s all happening at a speed that nobody can handle. Including our own computer systems.

The difference is, the computers are going to adapt. 1Password can raise its limits. Microsoft can add servers. That’s a bandwidth problem, and bandwidth can be fixed with money.

Human bandwidth can’t. Not without some science-fiction brain upgrade. That director can’t buy a second brain. Rocky can’t install another processor. I can’t either.

So what’s that director doing? He’s asking his assistant to keep his inbox clean. Which means there are things that technically come to him that he’s never going to see.

I imagine that’s what every CEO of a big company deals with. You can’t look at everything. You have to decide what you’re going to look at.

Maybe that’s not a new problem. But it’s a new problem for people who are used to knowing every single detail.

And what I’m hearing isn’t people fighting to hang on to the details. A staff member at that same nonprofit told me something like, “I don’t want to be involved in all that detail.” The director is saying the same thing with his inbox.

They’re asking to remove themselves.

I talked all this through with ChatGPT, and it said something that really got me. At some point the human in the loop becomes ceremonial.

Everybody says the safe way to use AI is to keep a human in the loop. AI does the work, a person checks it.

Except what happens when the AI can do a thousand things in the time it takes you to really check ten? You’re not checking anymore. You’re a rubber stamp.

I’m already there with code.

Systems I built a year or two ago, AI told me everything was great. Now I have newer AI look at them and it finds massive holes. Every time. It’s amazing, and a little terrifying.

That’s exactly what happened with Java and PHP. We all thought everything was fine. It wasn’t fine.

And I don’t understand the holes. I don’t understand the fixes either. So I have another model check the first model’s work. The answer to protecting us from AI is more AI.

So no, I’m not really in the loop on that code anymore. Not in any way that means anything.

But I did fix the 1Password problem. And I didn’t fix it by watching every request go by. I fixed it with one number. 250 a day.

That’s really the point. I can’t read everything. Rocky can’t. That director can’t. Nobody can. What we can do is set limits. How much a day. How many words. Who gets to see what. And then let it run.

I don’t think that’s giving up control. I think that’s what control looks like now.

Funny thing. That ChatGPT conversation where I worked all this out? I wrote about 1,500 words. It wrote about 10,000. Every answer started with some version of “that’s a fascinating point” and ended with another question to keep me going.

I loved it. And if I’d pasted the whole thing into an email to somebody, I’d be part of the problem.

The computers will get bigger pipes. I’m not getting a bigger brain.

I love you, Sage


I wrote this with help from ChatGPT and Claude. The stories and opinions are mine.