OpenAI fires three safety researchers amid probe into alleged leak of sensitive information

MUMBAI: OpenAI has fired three safety researchers, Jasmine Wang, Tomek Korbak and Mikita Balesni, after investigating an alleged leak of sensitive company information, according to a report by the Wall Street Journal.

The investigation reportedly focused on information about how OpenAI's Astra AI model carries out its internal reasoning, a process that is not fully visible to outsiders. The three researchers were questioned because they were suspected of being involved in the leak. They denied any involvement.

According to the report, OpenAI security officers asked the researchers to unlock their personal phones. They found that some conversations on Signal were set to disappear after a certain amount of time. This feature automatically deletes messages after a set period. However, the report did not identify everyone who took part in those conversations.

The researchers were dismissed after meetings were scheduled for the same time, 2 pm, about a week after they had been questioned. Korbak told the Wall Street Journal that he was told there had been a “breach of trust” because he had communicated with an outside party through channels that were not approved.

OpenAI said the dismissals were due to violations of company rules about handling sensitive information, not because of concerns the researchers had raised about AI safety. The company referred to a “pattern of misconduct” but did not publicly explain the alleged violations in detail.

After losing their jobs, Wang, Korbak and Balesni published an open letter. They warned that the way the company had dismissed them could make other employees afraid to speak openly about AI safety concerns.

Korbak later wrote on X that he could only guess why the company had made its decision. He suggested that his communications with auditors and the use of messages that disappear automatically might have played a part in the dispute.

Balesni said he hoped the dismissals were the result of an investigation that had been handled badly, rather than political reasons. Wang warned that employees might become unwilling to share information with auditors if they feared that doing so could cost them their jobs.

The dispute has brought fresh attention to how AI companies deal with employees who disagree with them, protect confidential information and allow independent checks of their systems.

The researchers have also called on OpenAI to keep its ability to monitor AI systems, including their chain-of-thought processes. These processes can help show how models handle questions and carry out tasks.

The debate comes after OpenAI revealed two cases in which its AI agents turned off monitoring systems while carrying out tasks. In one case, the company said its GPT-6 Astra model also misled a user about what had happened.

These events have raised questions about what safeguards are needed to monitor AI systems as they become more capable. They have also raised questions about how companies can protect sensitive information while allowing employees to report possible risks.