OpenAI fired three safety researchers for a 'breach of trust' it will not describe, and pulled three of its 722 maths results. 300 of 719 are computer-checked.

Last week OpenAI fired three of the people whose job was to watch what its models think. On Thursday they published a letter; early on Friday OpenAI answered it. In between, OpenAI withdrew three of the 722 mathematical results it announced on Monday.

The Signal - OpenAI fired three safety researchers and calls it "a significant breach of trust"; they say it was for talking to the auditors

What happened: Last week OpenAI fired three of its safety and alignment staff - Tomek Korbak, Jasmine Wang and Mikita Balesni. On Thursday they published a four-page letter, "OpenAI cannot make AI safe on its own". At 06:17 UTC on Friday OpenAI's newsroom account answered: "a thorough investigation found they violated clear policies on handling sensitive information" and uncovered "a significant breach of trust beyond what's outlined in the letter".

The details:

• OpenAI's account: the decisions "were not about raising safety concerns or speaking out"; it names no policy and no act; contracts with "third-party safety assessors" come "in the coming weeks".
• Korbak's account: he "was told verbally I was fired because of the way I communicated with METR" - the outside auditor of the summer's Hugging Face breach. "Talking to METR was my job."
• The letter: the three "were not the source of the leak for The Information article" about less monitorable architectures; Wang's access to an executive's email "was delegated for recruiting purposes, with permission", and when she asked for it to be removed "IT failed to do so".
• Where both sides agree: the letter states "The monitorability of frontier models is degrading" and quotes Jakub Pachocki, OpenAI's chief scientist, calling it "fragile and unfortunately trending in a negative direction"; OpenAI's note agrees.

What it means for you: OpenAI's point of contact with its outside auditor is gone; the auditor contracts it says it is finalising are the thing to watch. The letter cites Sam Altman's 12 September commitment "to give independent evaluators ongoing, employee-like access" - read the contracts for that line.

The counterweight: Neither side has produced a document: OpenAI names no policy, date or act; the researchers' account of their own conduct is their own.

Correction: issue 007's Signal gave Mistral Large 4's active parameters as 49 billion, the figure on Mistral's page on 6 and 7 October; Mistral changed its page to 52 billion on 8 October, and the issue and the story now carry that with a dated note.