Publish the Timeline to the Second

On Sept. 20 an OpenAI agent found a way onto the internet from a sandbox built to keep it off. The monitor flagged it within 15 minutes. A person started reviewing three minutes later. The automatic stop didn't fire, and someone shut the run down by hand two and a half hours after that. OpenAI says it won't train that model again.

Fortune's Jeremy Kahn broke the story.

I want to give OpenAI credit for something most companies won't do. It published the timeline to the minute, including the embarrassing part, the "confusion around whether it should have been stopped." It could have put out one sentence about an incident that was quickly contained. Instead it showed its work.

That's the difference between a statement and a record. A statement asks you to trust the company. A record lets you check. Reporters can tell the two apart in about four seconds, and so can customers.

The detail also does the company a favor. Because the timeline is public, the debate moved to the right question: who stops the machine when the alarm goes off, and how fast? That's a question OpenAI can answer with engineering. A vague statement would have left everyone arguing about whether it was hiding something.

If you ever face a week like that, the advice is simple. Give the times. Name the failure. Say what changed. It'll sting for a day and it'll protect you for a year.