Safety in Self-Evolving LLM Agent Systems: Threats, Amplification, and Case Studies

The paper presents a security analysis of self-evolving LLM agent systems that autonomously update their components, revealing a new threat landscape characterized by permanent adversarial influences and self-amplification across generations. It introduces the Module-Lifecycle Attack Surface (MLAS) matrix to assess vulnerabilities across 25 functional modules and lifecycle stages, finding that 17 are critically threatened without effective mitigation. The study highlights that evolution-native designs significantly increase the attack surface and persistence of threats, necessitating the development of evolution-aware security frameworks and formal verification methods for self-modifying AI systems.