New audit method reveals hidden flaws in AI behavior-control tests

Original: SteerCheck: Attribution Specificity and Alignment Leakage in Activation-Steering Audits

Source: arxiv.org ↗