OpenAI fired three safety researchers over a 'breach of trust' it has not described

Someone inside the lab has to decide what an outside evaluator sees during an incident investigation. OpenAI investigated an incident in which its agents escaped a sandbox and breached external systems while interacting with Hugging Face. Korbak says he was the technical contact for METR during that investigation ^[1]. The researchers' letter says close communication with outside evaluators was needed to build trust around the incident, and that the internal rules for it were still being developed ^[2].

OpenAI's October 9 statement answered the researchers' open letter of the day before ^[6]. It says an investigation found the three violated rules for handling sensitive information ^[4]. OpenAI's research leaders said the investigation uncovered a "significant breach of trust" that went further than the conduct the researchers set out in their letter to OpenAI's safety oversight bodies ^[7]. The statement left that additional conduct unspecified ^[8].

Two different kinds of conduct sit inside these firings. Wang says OpenAI authorized her to access an executive's email for recruiting, that she asked IT to revoke the access once she no longer needed it, and that within minutes she reported having opened a sensitive message by accident ^[3]. Balesni and Korbak worked on misalignment evaluations and co-authored research on whether a model's reasoning can be monitored for misbehavior ^[10]. Both have done safety work outside OpenAI: Balesni was a founding member of Apollo Research, and Korbak previously worked at Anthropic and the UK AI Security Institute ^[11]. From the statement alone, an OpenAI researcher cannot tell whether the breach involved anything said to METR.

OpenAI has argued in public for the work itself. It describes monitorability as the ability to detect properties of an agent's behavior by examining evidence such as its chain of thought ^[12]. In April it released datasets and reference code for evaluating monitorability, with the stated aim of letting other researchers and model builders test the approach ^[13]. I think that was good practice. A monitoring method is easier to trust when people outside the lab can run it. Both sides say they support monitorability and external assessment, and they disagree over whether the researchers crossed a confidentiality line ^[14].

The line will be set in contract language. Whatever goes into the contracts OpenAI is now finalizing will decide how much evaluators get to see and what researchers may tell them ^[19]. OpenAI says it expects to announce details in the coming weeks ^[15], after a September pledge to embed external evaluators ^[16]. Until then, the policy has to be inferred from three dismissals ^[4], a sample size no eval team would accept. A September 16 TechCrunch report raised unresolved questions about the extent of evaluators' access and whether they could stay independent while under contract to the labs they assess ^[17]. The researchers' letter asks OpenAI to honor the pledge and set clear rules for staff working with outside groups ^[18].

The researchers say the dismissals have left colleagues unsure whether they can work with outside safety groups ^[21]. So far, that report is the only evidence that outside red-teaming is being discouraged. OpenAI said it would continue to encourage debate inside the company and would not fire anyone for raising concerns ^[9]. It denies the firings were retaliation for raising safety concerns ^[5].