OpenAI Fires Safety Researchers, They Fire Back

Three former OpenAI safety researchers are fighting back, saying they were wrongly fired for doing exactly what they were hired to do: keep AI safe.

Jasmine Wang, Tomek Korbak, and Mikita Balesni published an open letter on October 8, denying they broke any rules or leaked sensitive information. They warn their sudden dismissals are creating a "chilling effect" inside the company, making it harder to spot and respond to risks in advanced AI models.

What OpenAI Claims

Last week, OpenAI launched an internal investigation. The company says the three researchers bypassed established procedures to "access and process sensitive research information," and then terminated their employment.

But the researchers tell a different story. They say their work required close collaboration with third-party security assessment agencies—especially after unprecedented security crises like the "Hugging Face agent breaking out of the sandbox incident." Building trust with outside experts wasn't optional; it was essential.

The Compliance Question

Wang, Korbak, and Balesni insist they stayed within their authority and followed current company regulations. They also say their communication with external experts—aimed at addressing declining visibility in Chain-of-Thought reasoning for new architectures—had compliance support from senior management and board members.

Wang specifically addressed one incident: she says she promptly reported an executive's email being mistakenly opened to IT and executives. Her dismissal, she argues, was unfounded.

A Deeper Tension

OpenAI's internal memo reiterates encouragement for raising safety concerns and denies retaliation. Yet this episode exposes a raw nerve at the heart of AI development: the clash between protecting commercial secrets and enabling transparent third-party evaluation.

As AI agent security incidents become more frequent, drawing the line between strict confidentiality and open safety audits is no longer a theoretical debate. It's an urgent governance challenge—one that will shape how large models evolve and how much the public can trust them.

Key Points:

• Three former OpenAI safety researchers deny violating company rules or leaking sensitive information.
• They say their collaboration with third-party security assessors was necessary and had management support.
• OpenAI claims they bypassed procedures to access sensitive research information.
• The dispute highlights tension between commercial confidentiality and independent safety audits in AI development.