OpenAI endorses the audit plan of three safety researchers it fired for what it calls leaks

OpenAI's response to the letter contains two positions that are hard to reconcile. The company said the terminations had no connection to the researchers' safety advocacy and that it agrees with the letter's recommendations. It also called the dismissals necessary because of breaches of trust ^[2]. One of those recommendations is to bring in independent safety auditors ^[1]. CryptoBriefing describes the general problem for labs: outside validation of safety work requires sharing information, and protecting proprietary research requires keeping it close ^[8].

Tomek Korbak's previous job involved exactly that kind of sharing. He was OpenAI's liaison to METR and Redwood Research, two external safety groups ^[5]. According to CryptoBriefing, an internal investigation reportedly found that the three had passed confidential materials to an outside AI safety organization ^[11]. The researchers say their conduct fit within their professional responsibilities ^[11]. The account does not name the organization that received the material or say what it contained.

The letter's technical warning is about chain-of-thought monitoring, in which safety researchers watch the intermediate reasoning a model produces before its output for signs it is doing something it shouldn't ^[1]. The authors want that monitoring work done jointly across the industry ^[1]. They also accuse OpenAI of putting profit ahead of safety ^[6].

The dates are close together. The letter went to the board and its safety committees seven days after the dismissals ^[9]. Add David Robinson's resignation as safety transparency lead ^[7] and four people left safety roles at OpenAI within a few days ^[14]. The last round was roughly two years ago, when the company fired researchers over similar violations in 2024 ^{[13] [15]}.

For anyone with money in OpenAI or a lab like it, two readings fit these facts. In one, the company enforced an ordinary confidentiality rule against staff who happened to work on safety, and its support for independent audits is sincere. In the other, the confidentiality rule is the real limit on outside oversight, and agreeing with the letter costs nothing until an auditor has written terms of access. I think the second fits the record better. OpenAI has put its enforcement behind confidentiality and only its words behind the audit plan, standing by the dismissals while agreeing with the letter ^[2].

A formal arrangement with independent auditors, setting out what they may see, would prove that view wrong. CryptoBriefing lists such an announcement among the next signals to watch ^[12].