One way to build large-scale autonomous systems is to develop open peer-to-peer architectures in which peers are not pre-engineered to work together and in which peers themselves determine the social norms that govern collective behaviour. A major practical limitation to such systems is security because the very openness of such systems negates most traditional security solutions. We propose a programme of research that addresses this problem by devising ways of attack detection and damage limitation that take advantage of social norms described by electronic institutions. We have analysed security issues of open peer-to-peer multi-agent systems and focused on probing attacks against confidentiality. We have proposed a framework and adapted an inference system, which shows the possibility of private information disclosure by an adversary. We shall suggest effective countermeasures in such systems and propose attack response techniques to limit possible damages.