OpenAI Researchers Warn of Lost AI Safety Tool

Two former OpenAI researchers are speaking out about a safety mechanism they say got lost in the shuffle after they were pushed out of the company. The tool in question involves monitoring how AI models reason through problems, a method some researchers consider a check on systems that otherwise operate as black boxes. Its disappearance, according to the people who built it, didn't come with much explanation.

The episode lands amid broader unease about how fast AI companies move and how little outside scrutiny their internal decisions get. OpenAI has cycled through multiple safety-focused departures over the past two years, each exit accompanied by warnings that commercial pressure is crowding out caution. This case is narrower: it's about a specific technical safeguard for reasoning models, not a sweeping indictment of the company's direction.

What makes it notable is who's raising the alarm. These aren't outside critics or policy advocates. They're the researchers who built the tool, fired, and now left to describe what they say got dropped once they were out the door. Their account raises questions about what else might fall through the cracks when safety teams turn over.