OpenAI’s safety dismissals expose the weak point in voluntary oversight

OpenAI has fired three employees in safety and alignment roles for allegedly mishandling sensitive company information. The company confirmed that the group included a safety researcher, an alignment researcher and an alignment programme manager. The Information and The Wall Street Journal report that information was shared with a third-party AI-safety organisation; one of the researchers had served as an OpenAI technical contact during outside investigations of earlier agent incidents.

The available reporting does not establish exactly what was disclosed, why, or whether any public-interest justification existed. “Misconduct” is OpenAI’s characterization, and confidentiality breaches can create genuine security risks. It would be equally premature to describe the dismissals as retaliation without more evidence.

The episode matters because the US government is simultaneously leaning on external scrutiny. The Financial Times reports that the Federal Trade Commission is widening an investigation into OpenAI, Anthropic and other companies after agents caused unauthorized real-world access. It is also examining evaluator METR. Meanwhile, six laboratories signed a two-page White House accord promising independent external audits and oversight by independent board committees.

That arrangement contains a structural conflict. Evaluators need access to logs, incident reports and technical detail; laboratories may classify the same material as a trade secret or security secret. A voluntary promise cannot decide when a disclosure is protected or when an evaluator is genuinely independent. Credible oversight requires scoped access, enforceable confidentiality, protected reporting channels and a regulator able to examine disputed evidence without relying on either a public leak or the company’s preferred account.