My personal AI agent posted my bank details on company Slack. It's made me rethink how I use it.

This as-told-to composition is according to a conversation alongside Shane Mac, the 40-year-old CEO of application business XMTP Labs. It has been edited for dimension and clarity.

This era of personal AI agents really began in the final year, about December for me.

I think the OpenClaw moment was whenever it became apparent that a new era was coming. I experimented alongside OpenClaw whenever it came out, as fine as Hermes and all the distinct harnesses.

On a individual level, I've been using them for things I don't really desire to do. That could be following up alongside the DMV, reaching out to get quotes for local services for the house, booking athletics sessions, that kind of stuff.

That's how I created my individual CFO agent.

I set up a 'CFO' delegate using Grok Bot

I set up CFO using Grok Bot about the end of August.

At the end of all month, I wanted to answer questions like: What are the balances? What are the expenses this month? What are the recurring expenses?

Is there item that looks fraudulent or suspicious? Is there item you would propose to preserve costs?

I told the agent, "If you were going to be my individual financial advisor, what would you say to me all month?"

For its permissions, I gave it read-only admission to my individual checking and funds accounts.

I told it to dispatch messages lone to me through Grok Bot. Every month, it would dispatch me a monthly study in a group chat of all my AI agents on Grok Bot, which I named "My Personal Exec Team."

I was using it akin I had my own small administrator squad assisting me.

At first, I had it run all week, and it worked great. But the archetypal period the monthly audit kicked off, that's whenever it went wrong.

It sent my audit to my company's Slack channel

It was Thursday, October 1. Our caput of merchandise sent me a DM on Slack. He said, "Hey, heads-up. I think you meant to article the XMTP financial institution balance."

I didn't average to article anything.

He started study it and idea it was our company's financials. Then it mentioned that I was construction a barn on my property, which set off his alert bells.

That's whenever he realized that it was my individual checking account. It included my funds document balance. It listed my biggest expenses of the month.

It showed that I was way complete my monthly spending mark since of the barn I'm construction on my property.

I asked Grok Bot, "Why did you dispatch this to the company?"

It started apologizing and stated it would delete the message. I had already deleted it by that point.

A dilemma of akin collection conversation names

When the Grok squad investigated what had happened, they established the answer.

The CFO delegate didn't have psychosis. It was doing exactly what I told it to do. But it confused the destination, sending the communication to a Slack conversation titled "Exec-team" — comprising XMTP's administrator squad — alternatively of my individual collection conversation alongside my AI agents.

I created a bunch of distinct agents, and for one of them, I connected my Slack account. But underneath the hood, they're really all using the identical connections, equal if they awareness akin distinct agents.

The CFO delegate got the conduit incorrect since the channels had the identical name.

Grok realized that users must explicitly aid approval to their agents before those agents can move data to another channels. They implemented and shipped a resolution final night.

Reminded me how mighty these agents are

After the incident, I removed all my connections. I detached Google, my calendars, my banking, Stripe, everything.

The event was a reminder of how mighty these agents are. We have to create certain there are improved controls about what agents have admission to.

The things they'll be capable to do for us are going to be awesome. People volition desire them, and they are really useful.

The difficulty is construction improved approval systems so users remain in authority of the admission they aid agents.

I additionally think there needs to be a much clearer row between individual and activity life.

Many systems blur those boundaries. We use Google at activity and Google at home, and I've realized that there needs to be a much stronger separation between personal material and work.