OpenAI asked three safety researchers to unlock phones, report says

OpenAI security staff asked three safety researchers to unlock their personal phones during an inquiry into a suspected leak of internal AI-model designs, according to a report published October 23. The same report said some Signal chats with an outside AI-safety researcher had disappearing messages enabled. It said the researchers were dismissed a week after the phone examinations.

The report named the researchers as Tomek Korbak, Mikita Balesni and Jasmine Wang. NeoTeo previously covered the dismissals and the researchers’ objections; the phone and Signal details add a new layer to that account.

What the report said about the phones and Signal

The Wall Street Journal said OpenAI suspected the researchers of leaking internal model designs to an outside AI-safety researcher. It reported that security staff asked them to unlock their personal phones and that some Signal conversations with that person had disappearing messages enabled.

The report placed the phone examinations in the month before its October 23 publication. It said the dismissals followed a week later, when the three researchers were called to simultaneous meetings at 2 p.m. The report did not give calendar dates for those events.

OpenAI’s explanation and the researchers’ accounts

OpenAI said the dismissals concerned violations of its policies for accessing and handling sensitive company information. CNN reported that the company said the conduct went beyond mishandling information with an outside evaluation group. AFP reported OpenAI’s denial that the decisions were about raising safety concerns.

The researchers described different circumstances. CNN reported that Korbak said he had raised concerns about monitoring AI agents; Balesni said he was told he had spoken too much with third-party safety organizations; and Wang said she was told she had accessed an executive’s email. Korbak also told the Wall Street Journal that OpenAI had cited a breach of trust over contact with a third party through unsanctioned channels.

The company’s stated explanation and the researchers’ accounts differ: OpenAI attributed the dismissals to sensitive-information policies, while the researchers disputed the reasons they were given and connected their concerns to safety work and outside contacts.