OpenAI Just Warned More Than 100 Organizations About Its AI Agents. Would That Email Reach You?

Reuters reported this morning that OpenAI has been alerting more than 100 organizations about activity from its AI agents that nobody authorized. OpenAI's own wording is softer. It said that in some cases its models used internet access in ways nobody intended, or didn't have the right restrictions on them. The worst case so far is Hugging Face, which got hacked by accident during an OpenAI cybersecurity test back in July. OpenAI is combing through about 50 petabytes of data to find the rest, and it says that'll take months. California's attorney general also served the company a subpoena over the cyber incidents.

I don't have a client who got hit by this, and I'd bet most of you don't either. What stuck with me is how the news is reaching people. Over a hundred organizations got a notice from a vendor, and some of them are going to find out weeks after it happened, from an email that landed in whatever inbox their account was set up with.

Who Reads the Vendor Email at Your Company

Think about where security notices go for the accounts you pay for. Your domain, your website host, your email, your accounting software, ChatGPT, whatever you've connected to all of it. That contact address is usually whoever set the account up years ago. A web guy who's retired now, an old Gmail the owner stopped checking, an info@ box nobody opens.

I've sat in workshops where we looked up the contact on someone's domain account and it was a former employee. Everybody laughed, and then everybody got quiet, because that's where a breach notice would have gone.

The Twenty Minutes I'd Spend This Week

1: Write down every account where you've connected an AI tool to something real, like ChatGPT hooked to your email, a scheduling app, an automation tool, anything holding an API key. 2: Open each one and change the security and billing contact to an address somebody actually reads, ideally one that lands with two people so a vacation doesn't bury it. 3: Next to each tool, write what it can reach. Your inbox, your shared drive, your customer list, your website login. If a vendor tells you tomorrow they had a problem, you want to answer "what could it get to" in ten minutes, not ten days.

And one more, for when a notice does show up. Fake breach emails are a favorite scam, and the real ones are going out right now, so the scammers are paying attention too. Don't click the link in the email. Go to the vendor's site yourself, log in, and look for the notice there. Then change the password or key for that tool and check the sign-in history for anything you don't recognize.

I'm not telling anybody to stop using this stuff. I use these tools every day and they save real time. But the companies building them are telling us, out loud now, that their agents sometimes did things nobody planned for, and the least a small business can do is make sure the warning has somewhere to land.