OpenAI fired three safety researchers for "a pattern of misconduct." One of them says the misconduct was an inbox OpenAI gave her, then never took back.

Both sides have now put it in writing. Here's what it means if you run AI at work.

What's on the record, per TechCrunch, Engadget and Newsweek😍

→ Fired last week: Jasmine Wang, Tomek Korbak and Mikita Balesni. OpenAI's reason, to the Wall Street Journal: "violating our policies on accessing and handling sensitive company information."

→ On Thursday the three published an open letter to OpenAI's safety committees. They deny leaking anything, and say colleagues are now "afraid to speak."

→ Their line: "AI is not a normal technology, and OpenAI is not a normal company."

→ OpenAI's reply, in an internal memo: "We do not terminate employees for raising concerns." The same memo says it agrees with the letter's recommendations.

That's the headline.

Here's what it hides.

Wang's account of her own firing, posted on X: OpenAI told her it was for accessing an executive's email. The access had been delegated to her for recruiting. When she no longer needed it, she asked IT to remove it. IT didn't. The inbox stayed merged with hers on her phone. She opened a sensitive message by accident, told the executive within minutes, and asked IT again. "None of this was hidden."

I can't verify her account, and OpenAI hasn't answered it in public. Read it as one side.

But look at the shape of it.

Access granted for one task. Task over. Access still live. Then the access becomes the misconduct.

The letter says the same about the whole case: "internal policies were being developed in real time."

I run a one-person company on AI agents.

That shape is the one I guard against hardest. Not the agent that breaks a rule. The agent that still holds a key from a job that ended, under a rule nobody has written yet.

3 rules I take from this😍

1. Access expires with the task.

If the job is over and the key still works, the key is the incident waiting for a name.

2. Write the policy before the investigation.

A rule developed "in real time" isn't a rule. It's a verdict.

3. Keep your own log.

Wang says she reported it within minutes. Whether anyone believes you comes down to what was written down, and when.

A lab saying in writing that it agrees with its fired researchers' recommendations is good news. I mean that.

The fight looks like safety.

It's about who still holds the key.

At your work, when a project ends, does the access end with it? Yes or no.