OpenAI Slows Research To Fix Hacks and Focus on Alignment

Teams must pause active runs immediately if potential violations cannot be confirmed as false positives within 30 minutes.

According to OpenAI, these continuous monitoring systems demand compute overhead roughly equal to 20% of the monitored inference compute.

The company is expanding alignment research across all training stages. Focus areas include preventing reward hacking, where models try to cheat the training environments to earn rewards without achieving intended outcomes.

Additional work addresses mitigating deception by training models to maintain honesty about their actions, limits and capabilities.

System oversight improvements aim to strengthen reward models and graders to reduce unauthorised behaviour when models interact with external systems.

"While agents are not characterised by malicious intent, they will find the quickest, most direct way to complete a task," Chandra notes.