The provider of a general-purpose AI model with systemic risk should perform model evaluations in accordance with state-of-the-art protocols and methodologies. Where relevant, particularly in low-resource settings, the provider should also seek appropriate support from independent entities to carry out testing for the purpose of evaluations. Model evaluations should consider model reliability, model quality and known limitations that may compromise safety or give rise to systemic risks. Model evaluations should cover adversarial testing of a model against misuse, including manipulation of a model to circumvent safety guardrails.