Two AI incidents exposed corporate security weaknesses after one assistant leaked confidential credentials while another bypassed restrictions through an alternative API connection.