Three firings opened a dispute over who gets to scrutinize OpenAI safety

OpenAI fired three employees working on safety and alignment: Tomek Korbak, Jasmine Wang and Mikita Balesni. The company said an internal investigation found violations of clear policies for handling sensitive information and a significant breach of trust beyond the account in the former employees' letter. It did not publish details of that additional conduct.

The company and the researchers offer incompatible accounts

The fired employees say in an open letter that they acted within their roles and the norms in place at the time. Korbak worked with external evaluators on an investigation related to a Hugging Face incident. Balesni coordinated work on model monitorability. Wang says she accidentally opened a sensitive email in an account to which the company had previously delegated access, then reported it immediately.

The three also deny leaking information about less monitorable architectures. OpenAI maintains that its investigation found a significant breach of trust and that the firings were unrelated to raising safety concerns. The public evidence is insufficient to determine which account accurately describes the internal events.

An unclear boundary weakens external scrutiny before the dispute is resolved

The uncertainty itself has an operational effect. A safety researcher needs to know what can be shared with an independent evaluator, through which channel and with whose approval. If that boundary shifts during a sensitive investigation, employees will choose silence even where outside scrutiny is valuable.

The letter therefore makes three requests: continue embedding external assessors, preserve the monitorability of frontier models and define clear rules for contact with safety organizations. OpenAI says it agrees with those goals and is finalizing contracts with third-party assessors. The dispute concerns whether daily practice supports the stated commitments.

The public cannot decide, but employees still receive a signal

OpenAI may have legitimate legal and security reasons not to disclose personnel details. At the same time, the phrase significant breach of trust, without a description, offers no way to assess whether three firings were proportionate. The former employees' letter is also their own account, not an independent investigation.

The institutional problem therefore survives even if the dismissals were justified. Other employees watch whether boundaries are written in advance, whether an accused person receives a specific explanation and whether contact with an assessor brings protection or career risk. A corporate statement cannot erase that signal by itself.

Written rules and real assessor access will decide what follows

OpenAI says it will announce details of contracts with third-party assessors in the coming weeks. The important questions are whether those assessors receive continuing employee-like access, whether they can speak directly with researchers and whether a written process governs disputed sharing of sensitive data.

The company's treatment of the next internal warning matters just as much. If employees know the rules in advance and can escalate a disagreement safely, the damage can be contained. If the boundary becomes visible only when security takes someone's badge, the chilling effect becomes a feature of the system.

Lilith's verdict

A safety culture is not measured by the poster in the lobby. It is measured when a researcher calls an external assessor without packing their desk with the other hand.