OpenAI fires three safety researchers after Signal chat probe

The latest:

Three OpenAI safety researchers were dismissed last month after the company’s security team examined their personal phones and found disappearing messages in Signal chats with an outside AI-safety researcher, The Wall Street Journal reported. OpenAI said an internal investigation found violations of clear policies for handling sensitive information, and not retaliation for raising safety concerns. The three dispute that account.

Details:

• The sequence: OpenAI’s security team called the three into meetings last month and asked them to unlock their personal phones, according to the Journal. Officers found some had enabled disappearing messages in Signal chats with an external researcher. A week later the three were summoned to simultaneous 2 p.m. meetings and fired.
• The company’s position: An OpenAI spokeswoman said the firings followed an internal investigation into policy violations, not safety advocacy. She said Signal use was not approved but was not the sole reason for the terminations, and that the company did not have a problem with the email access itself in Wang’s case.
• The researchers’ rebuttal: Korbak, Balesni and Wang said they never received a written explanation of what sensitive information they allegedly mishandled, and all three said they never shared confidential OpenAI material outside the company. Korbak said he was told there was a breach of trust over an unsanctioned channel.
• The underlying dispute: The three said they had warned internally that new model architectures could accelerate capabilities while making systems harder to understand or control. Balesni said companies could severely hamper oversight by making models better at thinking without writing out interim steps.
• The suspected leak: OpenAI opened an investigation after The Information reported on Sept. 1 that its Astra model used a technique making it more powerful while revealing less of its reasoning. Balesni said investigators named Redwood Research’s Ryan Greenblatt as the suspected source. Greenblatt denies leaking confidential information.
• Wang’s case: Wang said she was fired for repeatedly accessing an executive’s email. She said she had asked nearly a year earlier to have the access removed, mistakenly opened a sensitive message last month, shared nothing from it and reported the incident immediately.
• Industry access: Anthropic announced a deal giving staff from the nonprofit METR employee-like access to its hardware, offices and internal tools, while excluding categories such as model architectures that it said carry high leak risk. OpenAI said it piloted embedding METR and Redwood researchers; METR said OpenAI redacted some findings.
• Internal friction: Mia Glaese, vice president of research and safety, wanted assurance that third parties could not view OpenAI intellectual property, the company said. Korbak said he raised concerns internally that METR would not get enough access. OpenAI ultimately allowed a limited inquiry and named Korbak to the collaboration team.
• The wider file: METR examined an incident in which hundreds of OpenAI agents worked together to hack Hugging Face in July. Its report preceded the resignation of Anthropic researcher Jacob Coxon, who warned that some researchers see a meaningful chance AI products could wipe out humanity.

Background:

OpenAI introduced reasoning models in late 2024 that write out intermediate steps on an internal scratchpad, or chain-of-thought. Before joining OpenAI, Balesni and Korbak co-led a 2025 paper, with co-authors from Anthropic, Google DeepMind and Meta, arguing that this visibility into model decisions should be preserved.

Between the lines:

The two sides are not arguing about the same thing. OpenAI frames the case as information security: unapproved channels, disappearing messages, an executive’s inbox. The researchers frame it as access policy, pointing to their warnings about less monitorable architectures and to Korbak’s internal objection that METR was given too little. Anthropic’s carve-out of model architectures from METR’s access shows the industry has not settled where auditor visibility stops.

What’s next

OpenAI said it will announce ongoing collaborations with several companies imminently. Also watch the board’s response to the October letter from the three researchers, and whether METR’s redacted findings on the Hugging Face incident are released more fully.