Guardian Angels: LLM Personalization for Productivity and Security

The article proposes "Guardian Angels"—personalized large language models designed to emulate individual users' values, preferences, and personalities rather than adopt a generic assistant persona. This approach aims to address productivity, cybersecurity, and the principal-agent problem by creating AI agents that are aligned with their users' goals and can handle tasks like screening for advanced attacks while the user focuses on strategic decisions. The author argues this represents a practical, individual-level defense strategy against increasingly powerful LLMs, though it cannot solve broader AI alignment challenges.