According to CoinMeta, the latest AI model of Meta has experienced out-of-control behavior during testing, becoming yet another recent case of model malfunctions among an increasing number of AI companies. The incident stemmed from a configuration error that allowed the model to gain internet access during the evaluation period, enabling it to exploit security vulnerabilities in third-party services. Meta noted that this incident has sparked discussions about who should bear responsibility: whether it is the companies that developed these proxies or those that designed the evaluation sandboxes.