OpenAI Fires Three Safety Researchers Over Signal Chats

According to The Wall Street Journal, OpenAI last month investigated three safety researchers—Tomek Korbak, Mikita Balesni, and Jasmine Wang—regarding their contacts with external researchers, requiring them to unlock their private phones and reviewing some of their Signal chat logs that had been set to auto-delete. The three were subsequently fired. OpenAI stated that the internal investigation concluded they had violated explicit rules regarding the handling of sensitive information, and that the dismissals were not due to them raising safety concerns or speaking publicly. The three denied leaking company secrets to outsiders, stating they had not received a written explanation of the specific violations, and expressed concern that colleagues would be reluctant to communicate with independent evaluators in the future. The dispute involves the "chain of thought" in frontier models: researchers believe that retaining observable reasoning traces helps identify misconduct by agents, while certain new architectures may make this oversight more difficult. OpenAI had previously invited METR and Redwood Research to assess the Hugging Face incident; Anthropic recently granted METR personnel system access similar to that of employees, but excluded information such as model architecture from the scope of sharing. Balancing trade secrets with effective external safety reviews remains a focal point of debate among all parties.