Somebody Hit 30 AI Companies in Four Days to Steal Keys. You Have One Too.

Anthropic put out a report Thursday on how people are actually misusing AI, and most of the coverage went straight to the scary parts. Foreign spy crews. Malware that rewrites itself when it gets caught. All real, all worth reading. But there's a finding buried in there that hits a lot closer to a shop off Monroe Street than any of that, and almost nobody picked it up.

Attackers are now going after AI keys on purpose. One group in that report, a crew that used to make its money hitting hotels, switched over and went after around thirty AI companies in four days. They weren't trying to break the models. They were after production keys.

Why a key is worth taking

A key is a long password that lets one piece of software talk to an AI tool on your account. If you've got a chat bubble on your website, or a Zapier automation that writes your listing descriptions, or a plugin somebody installed to summarize the contact form, there's a key sitting behind it.

Two obvious reasons a thief wants it. They can sell it, because there's a market. And they can run their own work on it for free, which means your card gets the bill.

The third reason is the one that should actually bother you. Whatever gets done with that key looks like you did it. It's your account and your invoice, and the usage log has your name on it.

Where yours probably is

I ask this in workshops and the room goes quiet. Most owners don't know they have a key at all, because the person who set the thing up is a nephew, a contractor, or a web company they stopped talking to in 2024.

Four places to look. 1: any automation tool you pay for, Zapier or Make or n8n. 2: the website itself, especially anything with a chatbot or an AI form on it. 3: your email, search for "api key" and see what somebody sent you two years ago. 4: shared docs and spreadsheets, where people park a key so the whole team can grab it.

The website one is the worst of the bunch. If a key got dropped into the page code instead of the server, anybody can read it by right clicking and hitting view source. That isn't hacking. That's looking.

The twenty minute check

Log into whatever AI account your business pays for. OpenAI, Anthropic, Google, whoever it is. Find the API keys page. You'll get a list with names and a last-used date on each one.

Anything you don't recognize, delete it. Anything that hasn't been touched in a year, delete it. If a contractor or an employee who's gone ever had it, delete it and make a fresh one. Nothing breaks permanently. Whatever still needed that key will start erroring, and then you'll know exactly what was using it, which is a much better way to find out than reading a statement.

Then set a spending cap. Every one of these platforms lets you put a monthly ceiling on the account and hardly anybody turns it on. Set it a little above what you normally spend. If somebody's running your key at three in the morning, that cap is the difference between a fifty dollar surprise and a four thousand dollar one. Turn on the usage email alerts too while you're in there.

What changed this year is the math. Finding one exposed key used to cost a person an afternoon of poking around. Now it's a program that runs all night across thousands of sites and never gets bored or distracted, and it doesn't care how many employees you have.

None of this takes an IT guy. It's four screens inside an account you already pay for, and I'd rather you spend twenty minutes on it today than an hour on the phone with your bank later.