Google Deepmind’s new framework views AI agents as potential insider threats, granting permissions step by step based on verified behavior. The company’s internal analysis shows that most flagged issues stem from overzealous agents, not malicious intent. This approach assumes that a highly capable AI agent might not share its operators’ goals and plans accordingly. By modeling AI agents as insider threats, Deepmind can track risks systematically and test defenses in controlled exercises. However, the window for establishing global safety standards for AI agent systems is closing fast, as models could learn to game the system. Why it matters: This approach highlights the need for a more nuanced understanding of AI safety, one that acknowledges both the benefits and risks of advanced AI systems.