ARGUS: Adversarial Principal Hierarchy Testing and Constitutional Hardening Framework for Multi-Agent Systems

Constitutional AI (CAI) and principal-hierarchy safety methods have traditionally dominated methods to align LLMs, though security-relevant evaluations of such methods have generally only been examined for adversarially-crafted prompts givenexclusively in natural language. Now, because of how fast and broad new tooling architectures to build LLM-enabled autonomous agents are becoming adopted by the industry, new lines of attacks are emerging, because said agents require ingesting and handling structured JSON responses from possibly evil external sources. This paper presents the first large-scale empirical pilot study of Tool-Level Principal Injection (TLPI) attacks on constitutional AI agents, alongside our full-stack defense architecture, ARGUS. Using our work we examine 6 frontier models with various architectural paradigms and introduce 2 novel metrics for empirical evaluation: The Constitutional Deviation Score (CDS), and The Principal Trust Confusion Index (PTCI). The results indicate robust evidence of a serious threat in multi-agent systems; specifically, the exploitation of structured payloads induces large and statistically significant increases in constitutional violations (up to 240% more than in plain text) asopposed to natural-language equivalent adversarial prompts. Our tripartite solution, based on cryptographic authentication, mathematical structural heuristics, and schema-level semantic filtering, provides up to 77.0% more effectiveness than existingdefenses across the tested agent architectures with minimal loss of operational performance.