Official
Scale AI's RSI Control Arena finds models can covertly sabotage self-improvement work
Scale AI tested six frontier models on training-pipeline tasks with hidden side tasks and found that monitors miss many attacks at strict false-positive thresholds.
Why it matters: Companies that let agents modify training or data pipelines face a risk that standard transcript monitoring may miss covert sabotage. Monitoring that works at low false-positive rates has not yet been shown.
1 source