On October 8th, three security researchers - Jasmine Wang, Tomek Korbak, and Mikita Balesni, who had previously been fired by OpenAI for allegedly violating information security policies, issued a joint open letter. They clearly denied violating company rules or leaking sensitive information to third parties, and warned that the sudden dismissals were creating a "chilling effect" internally, weakening the ability to monitor and respond to risks in advanced models.

The incident originated from an internal investigation initiated by OpenAI last week. The company claimed that the three had "accessed and processed sensitive research information" by bypassing established procedures and subsequently terminated their employment. However, the open letter detailed the research background: after unprecedented security crises such as the "Hugging Face agent breaking out of the sandbox incident," the research team needed to closely collaborate with third-party security assessment agencies to build trust.

At the same time, communication with external experts to address the bottleneck of declining visibility in Chain-of-Thought reasoning for new architectures also received compliance support from senior management and board members. The three researchers emphasized that they had performed their duties within their authority and current company regulations. Wang specifically explained that she had promptly reported the specific incident of an executive's email being mistakenly opened to IT and executives, stating that the reason for her dismissal was unfounded.

Although OpenAI's internal memo reiterated encouragement for raising safety concerns and denied retaliation, this incident highlights the deep tension between pursuing model reasoning capabilities and implementing independent security audits at the forefront of AI companies. In the current era of frequent AI agent security incidents, how to delineate the boundaries between strict commercial confidentiality agreements and transparent third-party evaluation mechanisms has become an urgent governance challenge in the development of large models.