Leveraging large language models for automated access policies generation: an agentware approach

Large language models (LLMs) are probably the most popular type of foundation models (FMs) today. The growing adoption of conversational tools such as ChatGPT offers new solutions for software development challenges and opens new scenarios in which FMs can be employed to autonomously perform specialized tasks within a software architecture (the “agentware” paradigm). If properly trained or instructed, FMs can produce code even for highly specialized tasks, thus benefiting small companies which may be lacking specific domain experts. Integrating FMs in software design, on the other hand, introduces challenges that must be addressed. In this paper we present a software architecture employing agentware for automating XACML access control policy generation. We build on a prototype pipeline to propose an effective integration of agents into a software system. We will show that our design allows for optimal use of the generative power of FMs while tackling their intrinsic limitations.