OpenAI researchers fired over leak allegations deny wrongdoing

The dismissal of three researchers

OpenAI recently terminated three of its security researchers: Tomek Korbak, Mikita Balesni, and Jasmine Wang. The company stated that the dismissals followed an internal investigation into violations of rules regarding the handling of sensitive information.

According to a report by The Decoder, the company specifically identified the use of unauthorized Signal chats with external researchers as a factor. The security team reportedly suspected the three of sharing internal architecture details with an outside safety researcher.

The researchers have publicly denied these allegations of leaking confidential data. They argue that they were never provided with a written explanation of which specific pieces of information they were accused of mishandling.

Allegations of retaliation

The three former employees suggest their firing was actually linked to their internal work on making AI models more verifiable. They have been vocal about the risks of AI and have advocated for systems that allow for better oversight of how models arrive at decisions.

Specifically, the researchers worked on

Specific allegations against Wang

Jasmine Wang faced specific accusations regarding unauthorized access to a leadership executive's inbox. She stated that she had requested to have this access revoked nearly a year before the incident occurred.

Wang also reported that she accidentally opened a sensitive email last month. She maintained that she did not share the content of the email and reported the incident to the company immediately.

OpenAI stated that the act of accessing the inbox was not the primary issue in the termination. However, the researchers maintain that the company's investigation focused on their interactions with external parties.

Infrastructure and safety audits

The situation involves a broader debate over how OpenAI handles independent safety audits. During a recent on-site review, workers from METR and Redwood Research identified numerous security problems within the company's infrastructure.

OpenAI reportedly redacted nearly half of the final report from these external researchers. The company subsequently released its own report detailing a new security architecture that operates without internet access.

The company claimed that the identified errors were fixed, and they resumed training in early October. This move follows a period of internal scrutiny regarding how models handle sensitive data and internal architecture details.

The chilling effect on safety work

Tomek Korbak expressed concern that the dismissals will create a chilling effect for other employees. He fears that staff members will now be afraid to speak with external researchers or voice their internal safety concerns to management.

The three researchers submitted a letter to the OpenAI board requesting closer cooperation with external safety organizations. They argue that independent oversight is necessary to ensure the safety of advanced AI systems.

While OpenAI plans to announce several new collaborations soon, the friction over these dismissals remains a point of contention. The incident highlights the ongoing struggle between maintaining corporate security and fostering a transparent research environment.

External researcher denials

The investigation into potential leaks also involved Ryan Greenblatt, who was hired by METR to conduct an independent study. While OpenAI's security team suspected him of being a source for internal information, Greenblatt has denied sharing any confidential data with journalists.

The researchers involved in the investigation were reportedly looking for Signal chats between the employees and Greenblatt. This conflict underscores the tension between internal security protocols and the collaborative nature of independent safety research.

Infrastructure security findings

The external researchers from METR and Redwood Research reported finding many security issues within the infrastructure of OpenAI. These findings were significant enough that OpenAI chose to redact nearly half of the final report before it was made public.

In response to these findings, the company developed a new security architecture that lacks internet access. They reported that the necessary fixes were completed, allowing them to restart the training process in early October.