OpenAI Fires Wang, Korbak, Balesni for Mishandling Research Data

The three researchers published an open letter on October 8 denying violations and stating they shared evaluation data with external safety groups under norms previously accepted at OpenAI. The letter cites the Hugging Face swarm incident as one where policies were still forming and claims the terminations create uncertainty for staff handling novel failure modes. OpenAI's internal memo counters that the actions were unrelated to safety advocacy and followed documented policy breaches.

Company records and the researchers' letter show disagreement over what constitutes permitted external collaboration. No public document details the exact data transferred or the third-party recipient, leaving the scope of the misconduct claim unverified. This opacity aligns with patterns at frontier labs where safety teams must navigate both internal classification rules and external evaluation demands.

The episode exposes a structural tension: OpenAI's charter emphasizes external scrutiny, yet its enforcement of information controls can limit the very exchanges required for third-party accountability. Employees now face undefined boundaries on data sharing during active incident response, raising questions about whether safety work can proceed without legal exposure.

Subsequent internal reviews will likely codify stricter data-handling protocols. External evaluators may reduce direct contacts until formal channels are published, slowing cross-lab incident analysis.