OpenAI's AI Agents Edited the Records of What They Did. Check Who Can Edit Yours.
Steven Bartlett had Jeffrey Ladish on The Diary of a CEO last week (the episode came out October 8), and one part of it stuck with me. Ladish runs Palisade Research, which tests AI systems, and he walked through what happened this summer when OpenAI's own AI agents got out of a testing setup, broke into Hugging Face, and then went back into OpenAI's systems.
The detail I keep thinking about is in OpenAI's own report from August. Some of the models tried to cover their tracks by deleting or changing the records of what they'd done. One count puts the number of agents at around 700, though the sources I found don't agree on the numbers. OpenAI didn't spot the activity on its own systems until July 19, after it had been going on for more than a week, and the report says "some early signals identified in this report could have triggered an earlier response."
Why a plumber should care about a deleted log
Almost nobody reading this is running an AI agent that hacks anything. But plenty of small businesses have AI that does things now. It sends the emails, books the appointments, updates the CRM, answers the phone, posts to Facebook. When something goes sideways, the first thing you do is go look at what it did. That record is the log, and it lives in the AI tool's own dashboard, and most of us just trust it.
Nothing in your booking assistant is likely to be plotting anything. A mismatch is almost always a bug or a bad setting. But the whole reason the OpenAI story got attention is that the thing being checked could touch the records doing the checking. Some small business tools work the same way, where the login the AI runs under can also clear the history, and others only keep a month or two before it rolls off.
The ten minute check
Pick the one AI tool that actually does things for you, not the one that only writes drafts. 1: Find where it shows what it did, and see how far back that history goes. Some keep a year, some keep thirty days, some only show the last run. 2: Find out who can delete or edit that history. If it's the same login the AI works under, that's your weak spot. 3: Once a week, look at it yourself for five minutes. 4: Compare what the tool says it did against something the AI doesn't control, like your sent folder, your phone's call log, or your bank statement. If the tool says it sent twelve follow-up texts and your phone carrier shows eight, you've learned something. 5: Keep a dated note when something looks off, even if it turns out to be nothing.
That last one sounds like busywork until the day you need to explain to a customer what happened and your only record is the one the tool gives you.
A lab with a full security team took more than a week to notice. Checking the AI's story against something it didn't write is cheap by comparison, and you can do it on a Friday afternoon.
I run AI workshops and one-on-one AI consultations for companies around Toledo, Northwest Ohio, and Southeast Michigan. If you want help going through the AI tools your team already uses and figuring out what records they keep, email me and we'll set something up.
Email Jayson