From Thinker to Society: Security in Hierarchical Autonomy Evolution of AI Agents

Comprehensive AI Agent Security Analysis

This paper presents a systematic review of AI agent security through the HAE framework, which partitions agent capabilities into three levels: cognitive autonomy (L1), execution autonomy (L2), and collective autonomy (L3). The paper demonstrates that agent security risks are not isolated vulnerabilities but a systemic phenomenon that amplifies as autonomy expands. At L1, cognitive hijacking, IPI, and memory corruption disrupt reasoning and goal alignment. At L2, the confused deputy problem, tool abuse, and unsafe action chains introduce execution-level risks. At L3, malicious collusion, viral infection, and systemic collapse reveal the breakdown of monolithic safety assumptions in collective scenarios. Looking forward, agent security research must transition from fragmented defenses to systemic adversarial resilience. As agents penetrate real-world deployments spanning software supply chains, scientific laboratories, and social networks, security challenges demand breakthroughs across three fronts: (1) establishing contextualized security benchmarks that cover high risk scenarios including typosquatting attacks and laboratory jailbreaks; (2) developing neurosymbolic coordination mechanisms that construct unbypassable safety invariants through formal verification; and (3) building dynamic immune systems that leverage red team coevolution and decentralized reputation protocols to achieve adaptive defense. The central objective of agent security lies in establishing a trustworthy ecosystem where the release of autonomy and the imposition of safety constraints reach dynamic equilibrium. This necessitates deep collaboration among academia, industry, and regulatory bodies. Only then can agent technologies serve as a reliable force driving scientific and societal progress.