OpenAI Staff Were Reportedly Asked To Unlock Personal Phones Before Dismissals, WSJ Reports

OpenAI reportedly asked three artificial intelligence (AI) safety researchers to unlock their personal phones during an investigation into a suspected leak of confidential company information, before dismissing them a week later, The Wall Street Journal (WSJ) reports.

The researchers, Tomek Korbak, Mikita Balesni and Jasmine Wang, were investigated over suspected disclosures involving OpenAI's internal AI model designs.

The company said their dismissals followed an investigation into violations of its policies on handling sensitive information, while the three researchers deny sharing confidential company material with outsiders.

Investigators Examined Signal Messages

WSJ reported that OpenAI's security team summoned the three researchers in September and asked them to unlock their personal phones.

Investigators examined their communications on Signal, an encrypted messaging application, and found that some conversations with external AI safety researcher Ryan Greenblatt used disappearing messages.

OpenAI suspected the researchers had leaked information about its internal model designs to Greenblatt, who was then associated with Redwood Research, an AI safety organisation. Greenblatt denied sharing confidential information with reporters, telling the WSJ that he takes confidentiality seriously.

The researchers were called into separate meetings at the same time a week after the phone examinations and were dismissed, the newspaper reported. Korbak said he was told that his interactions with a third party through unauthorised channels constituted a breach of trust.

The three researchers said they had not received a written explanation identifying the specific confidential information they allegedly mishandled.

OpenAI maintained that the decisions were based on policy violations, not on the employees raising concerns about AI safety.

Dispute on AI Safety and Oversight

The investigation unfolded amid disagreements over how AI companies should develop models while allowing independent researchers to assess their behaviour.

The dismissed employees had advocated preserving the visibility of AI systems' internal reasoning processes, which researchers can use to investigate how models reach decisions.

Balesni had worked on proposals encouraging AI companies to avoid model designs that make internal reasoning more difficult to monitor. He said he shared a proposal with colleagues and discussed it with senior company figures.

He also said he shared a version with Greenblatt after receiving authorisation from OpenAI to speak with him.

Korbak had been involved in OpenAI's collaboration with the nonprofit Model Evaluation and Threat Research (METR), which examined an incident involving OpenAI AI agents and the code-sharing platform Hugging Face. The collaboration involved investigating the agents' behaviour and assessing potential safety risks.

OpenAI Defends Its Decision

An OpenAI spokesperson told the WSJ that the three researchers were dismissed after an internal investigation found they had violated company policies for handling sensitive information.

The spokesperson also said the decisions were not made because the employees raised safety concerns or spoke publicly about them.

Wang offered a separate account of her dismissal, saying she was told it involved accessing an executive's email. She said she had previously requested that her access be removed and had accidentally opened a sensitive email. She maintained that she did not share its contents.

The three researchers have expressed concern that their dismissals could discourage employees from communicating with independent safety organisations.

OpenAI, meanwhile, said it supports collaboration with external safety groups and indicated that further partnerships would be announced.