Static Analysis of AI Systems

Toolkits for Artificial Intelligence (AI) are increasingly being used in government and industry. If such systems have access to sensitive information, it is important to know their security properties. A toolkit may contain low-level flaws, such as buffer-overrun errors that allow an attacker to gain control of the host system. Or, there may be flaws in the rule base of a system implemented with the toolkit that allow unauthorized access to sensitive information. We propose the detailed study of the security properties of CLIPS, a widely used expert-system shell in two parts. The first part will be a detailed analysis of the source code of the system using static analysis tools and other methods to find low-level flaws. The second part will be a study of the CLIPS language and the exploration of static and dynamic approaches to create secure CLIPS programs. A promising approach is to model the CLIPS system as a weighted push down system, and to use model-checking techniques to implement a range of security analyses, including termination analysis and information flow. The results will include a design for a tool to be prototyped in Phase II, and guidelines on how to write a secure CLIPS specification.