Breach of Trust or Duty to Warn? The Dilemma Behind OpenAI's Dismissal of Its Safety Staff

Loyalty, Confidentiality and the Duty to Warn

Sit With the Dilemma First

Imagine you are a safety researcher at a frontier AI lab. Your job is to find out whether the systems your company builds could behave in ways their makers cannot control. You come to believe that something important is being missed.

Outside experts at an independent safety organisation could help you test that belief. Sharing internal material with them might sharpen the warning. It might also expose information your employer considers sensitive, perhaps for good reasons such as security, competition or the risk of a dangerous capability being copied.

The easy answer would be to call the researchers heroes and the company a censor or to call the company prudent and the researchers reckless. Neither fits what we actually know.

What Is Known and What Is Disputed

Agreed facts: OpenAI dismissed Mikita Balesni, Tomek Korbak and Jasmine Wang in the week before 1 October 2026. On 8 October they published an open letter.

OpenAI's account: They violated its policies on accessing and handling sensitive company information by sharing it with an external AI safety organisation, which it calls a significant breach of trust. It says the dismissals were not about raising safety concerns.

The researchers' account: They broke no policy, their work with outside safety experts was part of their role and the dismissals were pretextual. Korbak says he had raised concerns that the company was losing the ability to monitor AI agents' reasoning. They warn the move will make colleagues afraid to speak up. These accounts cannot both be fully true and no independent body has yet ruled.

The Competing Obligations

The researchers owe:

• Confidentiality to their employer and colleagues, who trusted them with sensitive material.
• Possibly a duty to warn the public and peers about risks no single company should judge alone.

The company owes:

• Protection of sensitive material and even application of its own rules.
• But also an awareness that dismissing safety staff sends a message to every remaining employee, whatever the stated reason.

A Framework That Helps Philosopher **Richard De

George set conditions for justified whistleblowing: the harm must be serious, internal channels should be tried first, the whistleblower should have evidence and disclosure should have a reasonable chance of preventing harm. Applied here, the key question is not whether the researchers were loyal, but whether the company offered a credible internal route** for safety disagreements and whether sharing with an outside safety body was a reasonable use of that route or a step around it.

A Defensible Position

The resolution I would defend: confidentiality should yield to the public interest only through protected, accountable channels and the company must be able to show its reasons were not retaliation. That means independent review of disputed dismissals of safety staff and agreed rules for sharing with outside auditors.

Something is sacrificed: the company gives up some control over its information and researchers give up the freedom to decide alone what to disclose.

Why This Matters for the Exam

GS4 asks when organisational loyalty must give way to public interest. This case shows that the answer depends on whether institutions give conscience a safe place to speak.