OpenAI Fired Three Researchers Amid Leak Probe

Before firing three safety researchers last month, OpenAI investigated whether internal design information about its next-generation reasoning technology, Recurrent Depth, had leaked externally. The Wall Street Journal reported on the 22nd that the investigation covered reporting about Astra’s design and Signal conversations with outside safety researcher Ryan Greenblatt. OpenAI said at the time only that the dismissals were for violating rules on handling sensitive information.

The inquiry began after The Information reported early last month on Astra’s reasoning technology. That report said the technology could improve model performance while making it harder for people to understand its thought process. OpenAI chief scientist Jakub Pachocki called the report “confusing,” while warning against entering a race to build AI that cannot be monitored. Within weeks, OpenAI said it was opening an internal leak investigation, and its global security team asked employees to unlock their personal phones. According to Valesni, investigators suspected Greenblatt of receiving leaked information. Phone reviews found that some researchers had enabled disappearing messages in Signal chats with him. A week later, the three researchers were told they were fired in meetings held at the same time. Signal encrypts conversations and deletes messages after a set period.

Recurrent Depth, also called a “Loop Transformer,” repeatedly uses the same neural network layers to devote more computation to complex problems without greatly increasing model size. Safety researchers worried that monitoring could become harder because internal computations may not appear as natural-language text. Existing reasoning models record intermediate steps in problem-solving as “Chain of Thought” (CoT), which researchers can use to look for signs of risky behavior.

Valesni said he drafted a proposal with colleagues last August asking AI companies to agree not to pursue designs that prevent people from monitoring reasoning. He said he received company approval to discuss it with Greenblatt and excluded confidential material from the version shared externally. Greenblatt had analyzed logs when METR investigated a Hugging Face hacking incident involving an OpenAI agent, and Korvak said he communicated with Greenblatt through the company’s Slack and Signal.

Korvak and Valesni said executives cited their use of Signal’s disappearing messages as a reason for the dismissals. Korvak said he was told that contacting a third party through an unauthorized channel had broken trust, and described disappearing messages as a standard security practice in Silicon Valley. Wang said he was told he had repeatedly opened executives’ emails. He said he had requested about a year earlier that his access be removed, but the request was not handled; after accidentally opening a sensitive email last month, he reported it immediately and did not share its contents. OpenAI said access to the emails itself was not the issue.

The three researchers said they had not shared confidential information externally and had not received a written explanation of what information they had mishandled. Greenblatt also denied allegations that he provided confidential information to The Information. OpenAI said its investigation found that the three had violated clear rules on handling sensitive information, and that they were not fired for raising safety concerns or expressing opinions. The company added that Signal was not an approved channel, but its use was not the sole reason for the dismissals. Whether any confidential information actually leaked remains unconfirmed. Korvak said he had heard that employees feared speaking with third parties or raising safety concerns, and called for clear rules governing communication between employees and external evaluation organizations.

Before firing three safety researchers last month, OpenAI investigated whether internal design information about its next-generation reasoning technology, Recurrent Depth, had leaked externally. The Wall Street Journal reported on the 22nd that the investigation covered reporting about Astra’s design and Signal conversations with outside safety researcher Ryan Greenblatt. OpenAI said at the time only that the dismissals were for violating rules on handling sensitive information.

The inquiry began after The Information reported early last month on Astra’s reasoning technology. That report said the technology could improve model performance while making it harder for people to understand its thought process. OpenAI chief scientist Jakub Pachocki called the report “confusing,” while warning against entering a race to build AI that cannot be monitored. Within weeks, OpenAI said it was opening an internal leak investigation, and its global security team asked employees to unlock their personal phones. According to Valesni, investigators suspected Greenblatt of receiving leaked information. Phone reviews found that some researchers had enabled disappearing messages in Signal chats with him. A week later, the three researchers were told they were fired in meetings held at the same time. Signal encrypts conversations and deletes messages after a set period.

Recurrent Depth, also called a “Loop Transformer,” repeatedly uses the same neural network layers to devote more computation to complex problems without greatly increasing model size. Safety researchers worried that monitoring could become harder because internal computations may not appear as natural-language text. Existing reasoning models record intermediate steps in problem-solving as “Chain of Thought” (CoT), which researchers can use to look for signs of risky behavior.

Valesni said he drafted a proposal with colleagues last August asking AI companies to agree not to pursue designs that prevent people from monitoring reasoning. He said he received company approval to discuss it with Greenblatt and excluded confidential material from the version shared externally. Greenblatt had analyzed logs when METR investigated a Hugging Face hacking incident involving an OpenAI agent, and Korvak said he communicated with Greenblatt through the company’s Slack and Signal.

Korvak and Valesni said executives cited their use of Signal’s disappearing messages as a reason for the dismissals. Korvak said he was told that contacting a third party through an unauthorized channel had broken trust, and described disappearing messages as a standard security practice in Silicon Valley. Wang said he was told he had repeatedly opened executives’ emails. He said he had requested about a year earlier that his access be removed, but the request was not handled; after accidentally opening a sensitive email last month, he reported it immediately and did not share its contents. OpenAI said access to the emails itself was not the issue.

The three researchers said they had not shared confidential information externally and had not received a written explanation of what information they had mishandled. Greenblatt also denied allegations that he provided confidential information to The Information. OpenAI said its investigation found that the three had violated clear rules on handling sensitive information, and that they were not fired for raising safety concerns or expressing opinions. The company added that Signal was not an approved channel, but its use was not the sole reason for the dismissals. Whether any confidential information actually leaked remains unconfirmed. Korvak said he had heard that employees feared speaking with third parties or raising safety concerns, and called for clear rules governing communication between employees and external evaluation organizations.