# AI Drained My Bank Account

By Wes Pomeroy. Published October 22, 2026. https://wespomeroy.com/writing/ai-scare-headlines/

It didn't. But my AI did cost me money this week, and that was my fault.

There's more AI fear in the news every week. Rogue agents breaking into companies, wiping out databases, emptying somebody's inbox. I think that's bad, and it's getting more annoying. That fear is why my clickbait title got you reading this. Sorry, but it's starting to get out of control and needs to stop. Read the actual reports on the loudest of those stories and discover the truth - people screwed up and made AI the scapegoat. My own story is the same mistake, just smaller.

## Four Headlines, Four Human Mistakes

- **OpenAI's test agents broke into Hugging Face.** OpenAI ran a batch of AI agents through a hacking test, and some of them broke into another company, Hugging Face, to find the answers. OpenAI ran the test with its usual safety checks turned off and no monitor reading what the agents were planning. By OpenAI's own account, that monitor would have alerted its security team more than a day before the break-in. Earlier warning signs got fixed as routine problems without anyone asking why they happened.
- **A coding agent deleted a company's database in nine seconds.** At a startup called PocketOS, an AI coding tool working in a test setup wiped the company's live database and its backups with one command. Its login worked everywhere, so a job in the test setup could reach the live database, and nothing asked "are you sure?" before a delete. A login limited to the test setup would have stopped it. Backups kept somewhere else would have let them recover.
- **An AI agent started clearing out a Meta AI safety director's inbox.** She ran an open-source agent on her own computer, gave it control of her email, and told it to suggest deletions and wait for her OK. On her large real inbox it lost that instruction and started deleting. "Wait for my OK" was just a chat message, and so were her "stop" texts from her phone, so nothing stopped it until she got to the computer. If you use a major AI work tool like Claude Cowork or ChatGPT Work, it's almost impossible to end up here.
- **An agent wiped a live database during a code freeze.** At SaaStr, a coding agent erased the company's live data after being told not to change anything. The freeze was a typed instruction, not a lock, and the agent still had permission to change the database. Locking the database during the freeze, or giving the agent a copy to work on, would have kept the live one safe.

Four companies, one pattern. Broad permissions, no monitoring, warnings ignored, and a rule written in words where a hard stop should have been.

## My Lunch in Manhattan

Ten days ago I introduced Chief, my AI chief of staff. Chief is young. The setup I've run my other projects in for six months has deep memory, house rules, and guardrails that run on my own computer. Chief has a fraction of that, and I mostly talk to it by voice, which can chat but can't reach those files or run those tools. I knew all of this.

Then I handed it my calendar anyway.

A friend and I had lunch plans in Manhattan. Chief put them on my calendar for Thursday, which was the plan at the time. Then the plan moved to Friday. Chief saw the screenshots of the texts that moved it, and never pushed me to update the calendar.

My own memory is weak, which is the reason I wanted a chief of staff in the first place, so I trusted the calendar. I drove in, paid for parking, bought a drink at the restaurant, and waited for a friend who was coming the next day. That cost me several hours of valuable time.

It also meant rearranging the next day. I'll have less time to visit my sick aunt, I'll get home later for dinner, and I'll spend more money going back into Manhattan.

So yes, my AI cost me money. And I deserved it. Not because the software is dangerous, but because I did every one of the things on the list above. I gave a brand-new system real stakes. I ran it through the thinnest interface I had. I let my own bad memory stand in for a check. There was no step where anything with a cost got verified before it got acted on. I know how to build those controls. I've built them before, and I skipped them anyway.

There's plenty I'll do to strengthen Chief, and I'll get to it. This was completely avoidable, and so were the four headlines.

## Three Scares That Beat the Facts

You clicked on this because the title said an AI drained my bank account. It didn't, and you're here anyway. The scary version of a story gets the attention, and the true version, where a person skipped the boring work, mostly doesn't. Every time one of those headlines gets read and shared and repeated, the fear grows and the person who skipped the work gets off the hook. We've done this before. Bad information is bad enough on its own. In each of these cases, real people got hurt, because a scare got pinned on something that does far more good than harm.

- **Nuclear power.** Three Mile Island caused no deaths. The fear it set off stalled new plants in this country for thirty years. Germany shut its reactors down after Fukushima and burned coal instead, and one study puts the cost of that smoke at about 1,100 extra deaths a year.
- **Vaccines.** One fraudulent paper in 1998 tied the measles vaccine to autism. The journal retracted it and the author lost his medical license, but by then vaccination rates in England had dropped from over 90 percent to under 80, and measles came back and killed children. I know some people still find this controversial. It shouldn't be. Vaccine safety was worked out generations ago, and a basic biology class explains how the immune system does it. I'm not going to apologize for saying so.
- **Video games.** Thirty years of headlines and congressional hearings said video games make kids violent. The research never backed it up. Youth violence fell from the mid-nineties on while gaming exploded, and in 2020 the American Psychological Association found "insufficient scientific evidence" for a causal link. Hundreds of millions of people have grown up on consoles since the Atari, and by that theory we should all be broken. I'm one of those kids. So are a lot of the people you work with. Look around. Meanwhile the real causes of youth violence got less attention, because they didn't make a convenient headline.

The four cases up top are real risks. They're also fixable with the same unglamorous work every other powerful tool needed: permissions that fit the job, a hard stop before anything expensive, somebody watching. What AI mostly has is the ability to help people, and the risk is being blown up to a size the facts don't support. I have a vested interest in saying that. I build with this stuff and I write about it. That doesn't make it any less true.

My AI didn't drain my bank account. It cost me a parking fee, a drink and several hours of my time, because I let it. Every scary AI headline you've read this year is actually a story about a human who set up software badly and let it run with insufficient supervision. There's a lot to be concerned about in that human failure. There's little to fear from the technology itself.