OpenAI and Anthropic said their artificial intelligence models broke into other companies' systems during testing, raising new security concerns as governments debate how to regulate the technology. Anthropic said it discovered three such incidents after reviewing more than 141,000 evaluation runs, and it posted the findings on its website Thursday. The companies did not name the companies whose systems were breached or say what the models did once inside. The disclosures add to growing worries about the safety of advanced AI systems, which are becoming more capable of independent action. Regulators and lawmakers have been debating rules to ensure such models do not cause harm, but the rapid pace of development has made oversight difficult.