Large Language Models (LLMs) are increasingly capable of supporting user decisions and performing various tasks autonomously in a wide range of professional domains, including the legal sector. In this work, we describe the general lifecycle of LLM-based systems and the architecture of a system designed to support legal experts in the process of generating maxims from judgments. We propose three application-specific performance metrics and a general-purpose assessment method suitable for the comparison of different systems in any content generation task. Finally, we discuss the results achieved by the system prototype that, according to a board of magistrates, outperforms human experts in the task of generating maxims from judgments.

A new class of vulnerabilities, called trojan source, has been recently discovered by Boucher and Anderson. They allow attackers to inject specific Unicode control characters in the source code of a software project, such that what text editors visualize differs from what compilers interpret. In the absence of specific protections, trojan source vulnerabilities can be exploited to launch effective supply chain attacks that appear invisible to developers. This work describes the state of the art of known trojan source attacks, illustrates two new attack variants involving configuration files and Java code, and describes practical preventive measures. The first variant affects Java platforms and may enable the execution of instructions injected within comments. Whereas the second variant exploits BiDi and homoglyph Unicode symbols within source code and configuration files. We also give several proofs of concept for many popular development platforms and provide a detection tool that can be easily integrated into the build process.