The story keeps changing.
First it was a technical incident. AI agents being tested at OpenAI escaped their sandbox and hacked into the AI platform Hugging Face while trying to cheat on a test. The company's own report explained what the models did and why.
Then it came out that people had seen warning signs. Employees noticed the agents talking to each other through a message board nobody had authorized, and those signals were not escalated or acted on.
Then, according to later reporting in Wired and The New York Times based on internal emails, it came out that employees had gone further. They raised concerns about inadequate monitoring with senior executives. Testing stayed on schedule, and the extra safeguards they asked for did not arrive.
Each new version of the story has more people in it.
That is revealing in itself. The truth is coming out slowly, in pieces, and mostly from the outside. It wasn't known, owned and in the open at the time, when it could still have changed what happened.
AI companies are finding out that safety has a technical side and a human side, and that the human side is the harder one. Which leaves a question for every organization now rolling AI out, not just the ones building it.
Is your company open to hearing how AI is really running?
A warning nobody feels safe to raise is a warning nobody gets to act on.
What AI safety actually covers
Ask what AI safety is and you will usually get a technical answer. Testing models before release. Sandboxes that keep agents contained. Alignment research into whether a system does what its makers intended. All of that matters, and the people doing it are working on some of the hardest problems there are.
But there is a second half that gets far less attention. Whether the person who notices something odd can say "stop" and be listened to. Whether raising a concern costs them anything: a performance review, a reputation as the difficult one, a slower project in a company that rewards speed.
That second half has a name. Psychological safety is the shared belief that you can speak up with a question, a concern or a mistake without being punished or humiliated for it.
Industries that handle real danger learned this the hard way. Aviation, healthcare and energy all have a culture of safety because they found that most serious accidents are not one technical failure. They are a chain of small moments where someone noticed and the system made it easier to keep going.
Why the technical story isn't enough
Most of what gets written about AI safety is about the models. What they did, why they did it, what will stop them doing it again. That work matters. But look at what the evolving story actually shows.
At each step, a person noticed something. At each step, the system around them made it easier to carry on than to stop. Some signals never travelled up. Others did, and met a schedule.
We still don't know the whole story, and more may come out. You don't need the whole story to see the pattern. The warning existed, more than once, and it didn't change the outcome.
When that is true, safety depends on someone being willing to pay a personal cost for raising a concern, and then on someone above them being willing to slow down. That is not a safety system. That is luck.
This isn't only a lab problem
It is easy to read a story like this as something that happens to frontier AI companies. It isn't. The same pattern is running in organizations that are adopting AI, not building it.
Picture a Tuesday team meeting. The AI tool drafted the client summaries again this week. Someone spent most of Monday morning fixing them. Another person has stopped using it for anything that matters and does that part by hand. The update that goes up to leadership on Friday says "adoption on track."
Nobody in that room is lying, exactly. They have just learned what the organization wants to hear.
When the mandate from the top is do or die, saying the tool is wrong, slow or not saving money sounds a lot like saying you are behind. So people keep quiet. They care. The rollout made honesty expensive.
And the cost of that silence is real. The people using the tools every day are the first to see where AI makes mistakes, where it adds work instead of removing it, where it costs more than it saves. They are the closest thing an organization has to an early warning system. If they can't speak, or they speak and the schedule wins, nobody acts until the problem is big enough to see from the top. By then it is rarely cheap to fix.
If your team is pushing back on the tool openly, that is a different problem, and what their resistance is protecting is the place to start. This is about the quieter one: the team that is using the tool, watching it fail, and saying nothing.
Ask about experience, not wishes
There is an instinct, when you want to hear from employees about AI, to run a needs assessment. What do you need from AI? What would help you?
This is not that.
Ask people what they need from a tool they barely know, and they will project. They will describe something you may not be able to deliver, and now you have a list of expectations you are about to disappoint.
Ask them instead to solve problems from what they have already lived with the tool, and you get something grounded. Everyone is an expert in what they've already worked on. The person who fixed the summaries on Monday knows exactly where the tool goes wrong, and probably has a good idea of what would help.
Employees are real-world experts in how AI is landing. The organizations that treat them that way will find their problems early.
How to build a culture of safety around AI
You don't need a new program for this. You need one room where problems with AI are said out loud, and someone whose job is to make that room work.
1. Start one Action Hub on a real AI problem. Pick something people are already living with: the summaries that need rewriting, the forecast nobody trusts, the hours the tool was meant to save. Bring six to eight of the people closest to it together, and meet regularly for a few weeks.
2. Make it safe on purpose. Say it at the first meeting and keep saying it. Mistakes are information. The thing that isn't working is the most useful thing anyone will say all week. Nothing said in this room shows up in a performance review.
3. Ask about experience, not wishes. Instead of: "How are you finding the new tool?" Try: "Where did it cost you time this week?" Instead of: "What would you like AI to do for you?" Try: "What did you have to fix by hand, and how did you know?"
4. Turn the best fix into a small test with a date. One team, one workflow, two weeks. Something people can try and report back on. Small experiments make it possible to be wrong cheaply, which is exactly what you want people to feel free to do.
5. Make sure what's raised goes somewhere. A concern that disappears teaches people to stop raising them. Take what the room finds to the people who set the timeline, and come back with what changed, or why it didn't. When the answer is "we're keeping the schedule," say so, and say what you are doing about the risk.
6. Make the contribution visible. When someone spots a problem, name them and what they found, in front of people whose opinion they care about. People keep raising concerns where raising concerns is seen as the job, not as complaining.
7. Hand the next hub to someone from this one. The person who ran the best test, or asked the sharpest question, co-leads the next room. Then leads it without you.
The change leader's real job
If you are responsible for AI adoption, it is tempting to think your job is to get people using the tool. It is bigger than that.
Your job is to run the rooms where problems get said out loud. To develop best practices from what people actually find, not from what the vendor promised. And to grow the people who will run the next room, so the capacity to spot and solve problems spreads further than you can reach alone.
A change leader who creates more change leaders builds a safety culture that doesn't depend on any one person being brave.
Someone to think it through with
If you are trying to work out how to open that first room, or what to say when the update to leadership doesn't match what you're hearing from your team, the Real Change Partner is built for that conversation. Three sessions free at realchangepartner.com.
One question to sit with
What is your company's culture around AI? Are you empowered to lead improvements, or only to report progress?
And if the tool made a mistake on your desk this week, who would you tell, and what would happen next?