Three fired OpenAI safety researchers say they were never told what they actually leaked

OpenAI's internal investigation into alleged leaks has led to high-profile firings and increased fear among employees about raising safety concerns.

Three OpenAI researchers specializing in AI safety—Tomek Korbak, Mikita Balesni, and Jasmine Wang—have been terminated for alleged policy violations, which they deny. They suspect their firings are linked to their advocacy for greater transparency in AI model reasoning and easier monitoring of AI systems. OpenAI stated the terminations were due to alleged policy violations, not the researchers' safety concerns, but the researchers claim they were not informed in writing about the specific leaked information. The dispute centers on the "chain of thought" process in AI models and the extent to which it should be accessible to external researchers. OpenAI's investigation reportedly focused on Signal chats with an external researcher, with one researcher also accused of unauthorized inbox access. The researchers expressed concern that these firings could deter employees from raising safety issues and discourage external safety auditors from collaborating with AI companies.

The firings highlight a growing tension between AI companies' desire for proprietary information and the critical need for independent safety oversight. This internal crackdown, coupled with external scrutiny of redacted audit reports, suggests a potential shift in how AI labs balance rapid development with robust safety validation. The researchers' claims of being unaware of the specific information leaked and the company's assertion that Signal was not the primary reason for termination create ambiguity, potentially weakening trust in internal investigations and future whistleblowing.

This situation could significantly impact the broader AI safety ecosystem by discouraging open communication and collaboration between AI labs and independent researchers. If employees fear reprisal for discussing model behavior or raising concerns, the pace of identifying and mitigating AI risks may slow. Furthermore, the incident raises questions about the effectiveness of communication tools and data handling policies in highly sensitive research environments, potentially leading other companies to adopt stricter controls or, conversely, to re-evaluate their approach to foster greater transparency.