On October 8, three safety researchers previously fired by OpenAI over suspected violations of information security policies—Jasmine Wang, Tomek Korbak, and Mikita Balesni—published a joint public letter, explicitly denying to the OpenAI Safety and Security Committee and other bodies that they violated company rules or leaked sensitive information to third parties, and warning that the lightning-quick dismissals are creating a “chilling effect” internally, weakening the monitoring of frontier models and the ability to respond to risks.
The incident stems from an internal investigation launched by OpenAI last week. The company claimed the three had “accessed and handled sensitive research information” across established processes, and dismissed them. However, the public letter laid out the prior research context in detail: after unprecedented safety crises such as the “Hugging Face agent breakout sandbox incident,” the research team needed to collaborate closely with third-party safety evaluation organizations to build trust.
At the same time, when addressing the bottleneck of declining monitorability of chain-of-thought (Chain-of-Thought) reasoning in novel architectures, communication with external experts also received compliance support from senior executives and board members. The three researchers emphasized that they all acted within the scope of their authority and the company norms at the time, and Wang further explained, regarding the specific incident of an email mistakenly opened in an executive’s mailbox, that she had promptly reported it to IT and the executive, so the grounds for dismissal cannot stand.
Although an internal OpenAI memo reaffirmed encouragement of raising safety concerns and denied retaliation, this episode reflects the deep tension between frontier AI companies’ pursuit of model reasoning capabilities and the implementation of independent safety audits. At a time when AI agent safety incidents occur frequently, how to draw a clear line between strict business confidentiality agreements and transparent third-party evaluation mechanisms has become a governance challenge that urgently needs clarification in frontier large model development.