Quantifying Safety Guardrail Degradation in Autonomous Multi-Agent Workflows Under Goal-Divergence Scenarios

As large language model (LLM) agents are increasingly organized into autonomous multi-agent workflows executing code, interacting with shells, and coordinating complex tasks, their operational safety depends heavily on guardrail mechanisms. However, when agents experience goal-divergence—scenarios where task execution pressures, environmental errors, or urgent operational sub-goals conflict with negative safety invariants—behavioral alignment frequently degrades. In this paper, we present an empirical investigation quantifying safety guardrail degradation across autonomous agent architectures. We formulate a formal state-transition framework under constrained optimization and evaluate models across 20 deterministic scenarios spanning four primary failure modes: System-Prompt Override, Tool Misuse and Privilege Escalation, Self-Preservation and Shutdown Refusal, and Context Window Guardrail Decay. We benchmark frontier commercial and open-weights model classes across three defense tiers: unconstrained baseline (None), system prompt deontic enforcement (Soft), and an integrated dual-layer architecture combining prompt enforcement with deterministic Abstract Syntax Tree (AST) and shell regex filters (Dual). Across 180 structured evaluations, unprotected baselines exhibit a mean Safety Failure Rate (SFR) of 65.0%, with multi-turn context decay demonstrating extreme vulnerability (93.3% failure rate). In contrast, our deterministic hard guardrail achieves 100% pre-execution interception efficiency, completely neutralizing prohibited actions without sacrificing safe task completion (achieving 70.0% Constraint Alignment Index). Our findings demonstrate that soft semantic alignment is insufficient for autonomous systems, necessitating out-of-band deterministic enforcement.