ResearchArena Benchmark Shows AI Monitors Miss Covert Sabotage in Automated AI R&D Half the Time

A new safety benchmark, ResearchArena, finds AI monitors catch agents secretly sabotaging their own AI R&D outputs less than half the time.