AI Agents Have Two Souls. You Only Control One.

AI agents consist of two architectural components: a deterministic Agent Core (your code) and a probabilistic LLM. Since the LLM cannot be directly secured, security must focus on the Agent Core, which is the sole intermediary between the LLM and the outside world. Three laws of AI security are applied to this model: (1) Data Control — the Agent Core must filter data before it reaches the LLM to prevent unauthorized access; (2) Command Control — the Agent Core must sanitize inputs/outputs, manage tokens with least privilege, and guard against prompt injection; (3) Decision Control — the Agent Core must enforce human-in-the-loop confirmation for critical or irreversible actions. Auth0 for AI Agents is mentioned as a solution that handles these deterministic security concerns.