AI safety fields face imminent automation

frontier labs are prioritizing the automation of technical safety research to support broader automated r&d goals. scalable oversight is the primary target due to its necessity for automating other fields, despite the circular dependency of requiring existing oversight to automate itself. mechanistic interpretability follows closely, benefiting from high feedback quality where techniques like linear probes allow for objective validation of model steering and feature verification. these shifts imply that new safety researchers should focus on areas with poor automation feedback or the underlying infrastructure for these automated pipelines.