OpenAI researcher departs as AGI readiness team dissolves

Another prominent safety researcher has just walked out the door at OpenAI.

Rosie Campbell, who spent nearly three and a half years tackling AI governance and safety policies at the company, announced her departure this week. Her exit follows the dissolution of OpenAI's AGI Readiness team and the recent departure of its leader, Miles Brundage.

Campbell joined in mid-2021 to spin up safety policies for OpenAI's API before moving to the Policy Research team. There, she worked on frontier policy questions like dangerous capability evaluations, digital sentience, and how to govern autonomous agentic systems. In an internal Slack note shared publicly, Campbell admitted that after the team's shutdown, she no longer saw a place internally to continue that work.

A blunt farewell to internal culture

Campbell noted that she felt unsettled by cultural shifts inside the company over the last year, especially the departure of foundational staff who helped define OpenAI's early mission.

Before leaving, she offered four pointed pieces of advice to her former colleagues. First, remember the goal is not merely building AGI, but ensuring it actually benefits humanity. Second, take seriously that current safety guardrails may fail against the vastly more powerful models expected this decade. Third, speak hard truths internally even when inconvenient. And finally, approach the work with appropriate gravity, reminding everyone that advanced AI is not just another consumer tech product.

For her next chapter, Campbell plans to take a brief hiatus to write and consult with independent researchers on critical governance gaps, AI moral patienthood, and evaluation limits.

Why it matters

Campbell's exit is another clear data point showing that long-term safety and speculative frontier governance are taking a backseat to commercial speed at OpenAI.

For the broader industry, dissolving the AGI Readiness team and losing senior policy voices accelerates an ongoing talent migration. Top researchers who understand the real risks of frontier models are relocating to think tanks, academia, and regulatory bodies. That external brain trust could help hold frontier labs accountable from the outside, which may prove more effective than fighting internal bureaucracy.

We think Campbell's parting note hits right at the central tension facing Silicon Valley today. When labs treat potentially transformative technology like conventional software releases, safety gets treated like a feature backlog rather than an existential safeguard. If current evaluation approaches can't keep pace with upcoming agentic systems, running ahead without dedicated internal readiness teams looks less like bold ambition and more like a massive unhedged risk.