Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research
arXiv — cs.AI · Keshav Shenoy, Anthony Zang, Eric Gan, Jacqueline Tay, Amy Ngo, Sebastian Prasanna, Julian Stastny, Vivek Hebbar · Published on Readzer
Summary
A study using a red-team/blue-team game found that human auditors failed to distinguish sabotaged from benign AI safety research, achieving an AUROC of 0.59 (not significantly better than chance) and fully identifying sabotage in only 17% of cases. In contrast, the best frontier models outperformed humans (AUROC 0.73, 32% full identification) despite not being allowed to run experiments. The results suggest that research sabotage by expert-level agents is difficult to detect with current auditing methods.
This is a reviewed summary. Read the complete article at the original source.
Why it matters by industry
- Artificial intelligence
This study directly addresses AI safety and alignment, highlighting vulnerabilities in detecting sabotage within AI R&D workflows, which is critical for the development of trustworthy autonomous agents and reasoning systems.