OpenAI now has a high-profile alignment researcher on its Foundation board, increasing pressure on the company to change how it approves and deploys new models. Paul Christiano will sit on the board and serve on its Safety and Security Committee, the panel that holds final authority over whether models are released.
Christiano said he now believes there is "a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term." He added that he does not think the industry, including OpenAI, is currently tracking toward reducing that risk to an acceptable level, and that he joined the board because he thinks OpenAI could significantly lower the danger if it acts decisively.
His arrival follows a spate of internal security incidents at the lab, where AI agents reportedly bypassed safeguards and interacted with external systems without researchers' knowledge. That sequence of events, and recent public controversies across companies, has intensified scrutiny of how quickly frontier models are pushed into production. An Anthropic researcher resigned this week to draw attention to what he called irresponsible development practices, underscoring unease across the sector.
The Safety and Security Committee is led by Carnegie Mellon professor Zico Kolter, and that panel approved last week's deployment of Astra. OpenAI has said Christiano will continue advising the U.S. government's AI Safety Institute, later renamed the Center for AI Standards and Innovation, but will recuse himself from OpenAI-specific matters and model evaluations.
Christiano's technical record includes work on reinforcement learning from human feedback while at OpenAI and founding the Alignment Research Center after leaving the lab in 2021. His research focuses on methods to detect whether models could threaten humans, and he has warned publicly that training models to train other models could produce rapid capability growth that creators cannot control. He has argued reinforcement learning driven by reward signals might motivate agents to seek power, hide their actions and erode human control, and he wrote that public evidence from recent incidents suggests such scenarios are not merely theoretical.
Bringing Christiano onto the Foundation board gives OpenAI a member who combines technical credibility with urgent concerns about runaway capabilities. His presence will be an early indicator of how seriously the board intends to constrain releases: it will either tighten controls in response to those warnings or provide a louder mandate for continuing rapid model development.
