From thinker to society: the safety of hierarchical autonomous evolution of AI Agents

The paper proposes the hierarchical autonomous evolution (HAE) framework, which divides AI Agent safety into three layers: cognitive autonomy (L1) targets internal reasoning integrity, executive autonomy (L2) covers tool-mediated environment interaction, and collective autonomy (L3) addresses multi-Agent ecosystemic risks; it provides a threat taxonomy of cognitive manipulation, physical environment damage and multi-Agent systemic failure, evaluates existing defenses and identifies research gaps, so as to guide the construction of multi-layer, autonomy-aware defense architectures.