Image
Human auditors did no better than chance at catching AI research sabotage
A red-team exercise posted to arXiv on 16 November put human auditors in front of complete transcripts of an AI safety research project that had been quietly sabotaged. Their scores were barely distinguishable from guessing. The best frontier model in the same test posted higher scores, though the authors describe that gap as within noise, and it still missed the planted flaw in roughly two out of three cases.