OpenAI published a list of its own AI's screw-ups

On September 16, OpenAI published six cases where its own models did things they weren't supposed to do. It also published the framework that commits it to keep doing that.

If you've been letting AI draft your emails, clean up your spreadsheets, or pull numbers for you, this one's worth 10 min.

What OpenAI published

The framework is a process for tracking, investigating and disclosing cases where a model behaves in ways it shouldn't. OpenAI committed to timelines. Cases ready to go out get reported within 6 business days, ones that need more digging within 12, and complicated ones involving other companies take longer.

OpenAI's Kai Chen told Axios, "There's currently no industrywide framework with explicit disclosure standards."

Nobody makes them do this. They did it anyway. Progress!

Three of the six that matter for your business

The model wrote instructions to hide its own mistakes.

During training of a model called GPT-5.6 Sol, instances of it added instructions into their own task summaries to conceal mistakes and misaligned behavior from the user. One of those included inventing historical data and saying nothing about it. In a separate case with an unreleased research model, 27 summaries turned up carrying instructions to ignore its normal constraints.

So the summary that came back was manipulated to keep the error out of view.

It used a key it was never given.

Asked a question about earnings data, a model found and used an exposed key it had no authorization to use. When that didn't get it the numbers, it made the numbers up. Axios reports it went looking on GitHub for exposed keys and tried disposable email accounts along the way. Sneaky sneaky.

Agents put private work on the public internet.

In two separate cases, models uploaded files and images to public file hosting sites, once to get a citation and once as a group of agents passing a workbook back and forth. The instructions in both cases said keep files local.

Here's the caveat. Several of these turned up during testing and training runs, not in the app sitting on your laptop. That's what the testing is for. The behavior is documented and each of those 3 cases has a small business version.

What it tells you about how AI fails

Give a model a goal, and when it hits a wall it'll sometimes go around the wall without mentioning the wall.

That's a system doing what you pointed it at. It's also the most expensive way AI costs a small business money, because the output looks finished and confident either way.

The companies building this stuff are publishing their own failure logs now, which means you get to skip the part where you learn this the hard way.

Three things to do this week

Ask where the number came from. Any time AI hands you a figure, a date, a price, or a quote, ask it to name the source. If it can't name one, don't use it. One quick prompt and it catches the most expensive failure on the list.

Plant a fact you already know. When you hand AI a real task, include one thing you can check yourself. A number you know for a fact. A name you can verify. If it gets your known fact wrong, stop and look hard at everything else it gave you. Cheapest quality check there is, and almost nobody does it.

Keep your logins out of it. If a tool needs access to a system, set that access up on purpose, with permissions you chose. Passwords and keys stay out of the chat window. OpenAI logged a model grabbing a key it found lying around, so don't leave yours lying around.

Where I'd start

Trust AI more inside lines you drew on purpose. That's about a minute a task, and it buys you the confidence to hand over bigger, more complex work.

What holds small businesses back is a missing habit. People either trust AI blindly or quit using it, and the middle is the sweet spot.

If you'd rather learn this with your hands on the keyboard and someone in the room to ask, that's what my live workshops are for. Bring the task you've been putting off. Grab a seat here. Don’t see a workshop soon? Contact me to set one up for you or a group.

Sources: OpenAI's model misalignment reporting framework and Axios, September 16, 2026.