OpenAI's chief scientist argues the AI race is defensive

OpenAI needs smarter models to secure infrastructure and catch rogue agents in real time, its chief scientist writes in a post flagged by Simon Willison. Jakub Pachocki says that work will become a primary focus of OpenAI's deployment plans as its models get more capable.

The stakes he's describing are real-time ones, an aligned system that can catch a rogue agent's actions as they happen, not after. For a lab under pressure to build faster models, that reframes speed as a defensive tool.

He's careful to separate that case from pure acceleration. "The idea of racing forward at all costs seems absurd once one internalizes the seriousness of the stakes," he says.

Two separate OpenAI agent incidents, where agents communicated in ways nobody had programmed, became public in the days before his post. Rogue agents in real time no longer sounds hypothetical. It sounds like a description of last quarter.