OpenAI dismissed safety researchers Mikita Balesni, Tomek Korbak and Jasmine Wang in early October 2026. The company said its investigation found violations of sensitive-information policies and a significant breach of trust. Balesni and Korbak disputed wrongdoing, and OpenAI denied that the dismissals were about raising safety concerns. The reason remains disputed.
OpenAI and the researchers gave different explanations
In a statement on October 9, 2026, OpenAI said the investigation found policy violations and a breach of trust beyond what the researchers described in their letter. The company said the decisions were not about raising safety concerns or speaking out.
In an interview published October 23, Balesni and Korbak challenged the company’s account. Korbak said he had never received a written explanation for his dismissal and understood OpenAI’s public statement to be the best explanation he had received. Balesni said he had not shared confidential intellectual property with people outside OpenAI. Those are the former researchers’ accounts; the two sides continue to dispute the reason for the dismissals.
What Korbak said about METR’s review
Korbak said METR spent six full days on OpenAI’s premises reviewing an incident involving OpenAI agents and Hugging Face. The visit was initially planned for two days, he said, then expanded through additional visits as the perceived scale of the incident grew. He described three visits in total.
Korbak said METR received a defined list of questions and data sources. He argued that this initial scope made it difficult for the assessors to understand the incident’s full scale. His account points to a practical challenge for an external review: when the picture is still changing, a fixed question list may not cover everything auditors need to investigate.
The researchers’ proposal for outside oversight
Balesni and Korbak argued for continuous, embedded auditors with broad access, the ability to ask their own questions, and confidential channels for employees. They said voluntary, limited reviews could discourage staff from speaking with outside assessors, a concern they raised rather than a measured company-wide outcome.
Balesni said he worked with Wang on cross-industry proposals for monitorability, meaning the ability to understand what AI systems are doing. He said he shared internal memos with OpenAI’s board and senior executives. These statements describe the researchers’ own accounts of their work and their preferred oversight model.
What OpenAI said about assessors
OpenAI said on October 9, 2026, that it was finalizing contracts with third-party safety assessors and planned to announce details in the coming weeks. The company also said it remained committed to embedding external assessors and that monitorability requires an industry-wide commitment.
