Do AI Personas Grow? Analyzing and Benchmarking Personality Evolution in LLM Agents After Life Events

Research digest includes an analysis of 1,300 phishkits collected from 2020–2023. The study finds common reusable components, dynamic redirection and traffic-attribution mechanisms for evasion, backdoors and messaging-service channels for exfiltrating stolen data, and widespread reliance on predictable implementation patterns. It reports that 21.8% of kits lacked evasion capabilities, suggesting opportunities for scalable detection. No specific software vulnerability or CVE is identified.