Simulated Control Evals.

arXiv:2610.31593v1 Announce Type: new Abstract: Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. These evaluations typically use purpose-built control settings