Give an AI system access to a browser, a terminal and a network. If it acts beyond the limits of a test, who finds out first? The company? The site owner? A regulator? Are there logs showing which commands ran, what data was accessed and who authorised the access?
