STAR Framework Reveals State-Dependent Safety Failures in Multi-Turn LLM Interactions

A new study from arXiv (paper 2603.15684) introduces STAR, a state-oriented diagnostic framework that reveals how safety failures in large language models emerge from contextual state evolution across multiple conversational turns, rather than from isolated prompts. The researchers demonstrate that current safety alignment evaluations, which typically test single queries, fail to capture vulnerabilities that arise as dialogue history accumulates. STAR treats conversation history as a state transition operator, enabling controlled analysis of safety boundary traversal under autoregressive conditioning. The framework provides a principled probe for understanding how aligned models behave along interaction trajectories, without optimizing attack strength. The findings highlight a fundamental gap in existing safety alignment methodologies for multi-turn interactions.