Guardian Angels: LLM Personalization for Productivity and Security

This article proposes 'Guardian Angels'—personalized LLMs that emulate individual user values and preferences to amplify rather than replace human decision-making—as a solution to cybersecurity, productivity, and principal-agent alignment challenges as LLMs scale globally. The approach combines dynamic evaluation, active learning, and continuous adaptation to create AI agents that are both more productive and more trustworthy than generic chatbots, while providing defense-in-depth against sophisticated attacks like synthetic media manipulation and social engineering. For IT organizations, this represents a fundamental shift from passive, one-size-fits-all AI deployment to active, personalized AI systems that require new architectural patterns, security protocols, and governance models to manage at enterprise scale.