The AI models were prone to safety failures and discrimination, a new study found.