OpenAI stands by firing three safety researchers

Ignore the personnel drama. The question underneath is who gets to check OpenAI's homework.

OpenAI is standing by its decision to fire three members of its safety team. The three have answered with a four-page letter, and both sides say they want outside auditors in the building.

OpenAI says breach of trust, the three say chilled dissent

The Wall Street Journal first reported the dismissals on Oct. 1, according to Fortune. The three are Tomek Korbak, Mikita Balesni and Jasmine Wang.

In its statement, quoted by Engadget, OpenAI said they violated "our policies on accessing and handling sensitive company information." Fortune reports OpenAI later said its investigation found a serious breach of trust beyond what the letter describes, and that it stands by the decision. An internal memo shared in part with Fortune says the firings were not about raising safety concerns.

The letter is addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. It says Korbak was the technical point of contact for METR, the outside evaluator, in the Hugging Face incident investigation, and that Balesni was involved in it too.

The three deny being the source of a leak to The Information about what the letter calls "supposed new, less monitorable architectures." They say Balesni's communication with outside parties was done in coordination with board members and the C-suite. They say Wang's access to an executive's email was delegated for recruiting, that she asked for it to be removed, and that IT never did it.

Korbak told Fortune he was told verbally he was fired for "the way I communicated with METR." OpenAI has not disclosed its specific reasons.

Here's the thing: the auditors are the story

The letter's demands are the part that touches you.

First, it asks OpenAI to honor its public commitment to embed third-party safety auditors, and to follow through on what it describes as Sam Altman's Sept. 12 commitment to give independent evaluators "ongoing, employee-like access." The three worry their firing could be used to end the METR work or narrow auditors' access.

Second, it asks OpenAI to preserve chain-of-thought monitorability, the ability to read a model's reasoning to catch bad behavior. The letter says it bluntly:

"The monitorability of frontier models is degrading."

OpenAI's position, per the memo Fortune saw, is that it remains committed to third-party safety assessors and will announce finalized contracts "in the coming weeks."

So both sides say they want auditors. Only one side gets to sign the contracts.

What it means if you build on OpenAI

You cannot audit a frontier lab yourself. You rely on people like METR, with real access, to check what the system cards claim. That is the whole value of red teaming by outsiders: it is not the lab grading itself.

If the auditors' scope shrinks, every safety claim you pass on to your own customers gets weaker, whether or not anything changes in the model.

Ask your account rep a plain question: which outside evaluators have access to the models you use, and what can they see?

Watch the contracts

OpenAI says the assessor contracts are coming in weeks. Read them when they land. Look for METR's name, the scope and whether access is ongoing or one-off.

Then decide how much of OpenAI's safety story you repeat.