Disputed dismissals at OpenAI raise questions about rules for safety work

Jasmine Wang, Tomek Korbak and Mikita Balesni, safety researchers dismissed by OpenAI last week according to TechCrunch, have published an open letter challenging the reasons given by the company. They deny handling sensitive information outside established procedures and say their removal risks discouraging their former colleagues from raising safety concerns.

The letter is addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. The researchers describe the ability to work with outside experts, without fear and under clearly defined procedures, as an essential condition for their work. They describe a cultural shift in which previously accepted behavior allegedly became grounds for dismissal. This description remains their account of events.

OpenAI maintains a different interpretation. A spokesperson told TechCrunch that an investigation had uncovered repeated breaches of rules governing the handling of research information, beyond sharing with an external evaluation group. The company also provided the outlet with an internal memo attributed to a research leader. It acknowledges the researchers’ contributions and states that their dismissals were not punishment for expressing safety concerns.

The researchers also dispute their involvement in a leak to The Information concerning model architectures whose reasoning would reportedly be harder to monitor. To explain their external communications, their letter cites, among other things, the investigation of an incident involving agents and Hugging Face. They say internal procedures were being developed in real time at that point. These details constitute their justification; the article does not establish whether their actions complied with the applicable rules.

The disagreement therefore concerns procedures as much as intentions. TechCrunch notes that OpenAI did not directly answer its questions about the specific rules breached, the circumstances of the dismissals or protections for employees working with outside evaluators. The available excerpt does not contain the investigation evidence needed to adjudicate between these accounts.

From an analytical perspective, the issue is the clarity of the limits placed on safety work. External collaboration may require technical exchanges, while the company cites the protection of sensitive information. Explicit rules would make this boundary easier to assess. The researchers call for preserving a culture of dialogue and honoring commitments concerning external auditors and the monitorability of models. According to the memo relayed by TechCrunch, OpenAI endorses these recommendations despite the disagreement over the dismissals.