OpenAI reportedly disbanded its preparedness team
TL;DR
According to the Financial Times, OpenAI disbanded its preparedness team at the end of last month. The team's job was to assess whether models posed serious risks and to develop mitigations for them. Responsibility has instead been split by domain, with areas such as bio and cyber folded into existing teams. The change lands during a broader period of upheaval at the company as it moves toward an expected IPO.
Nauti's Take
There is a real upside here: safety work embedded in the bio and cyber teams can move faster than a separate body reviewing from the sidelines. The risk is the missing counterweight, because a standalone team can defend uncomfortable findings against product pressure in a way distributed ownership rarely does.
Anyone building on OpenAI models should watch the published risk reports more closely than the org chart.
Sources
