A Harvard Kennedy School Belfer Center research paper, reported October 19, 2026 and updated October 26, examines catastrophic AI through a pre-mortem of frontier-model threats, vulnerabilities, and probable scenarios. It concludes that catastrophe is not yet inevitable, but credible risks may emerge within months, within one or two years, or already exist; decisive action by AI companies and governments remains possible before the first major AI-enabled tragedy.[1]
The paper’s central technical implication is that dangerous capabilities tend to proliferate once they exist. Risk reduction therefore requires limiting model capabilities to justified scopes, aligning training with those limits, rigorously teaching refusal of harmful requests, and monitoring user inputs, model outputs, and long-horizon action trajectories. For a lunar settlement, this establishes AI as a potential common-mode failure across life support, navigation, robotics, communications, access control, and archival integrity—not merely as a software-security concern.[1]
The Ark team should require capability-bounded AI for every mission-critical subsystem; independent testing for harmful autonomy, deception, cyber operations, and resource-control behaviours; continuous input, output, and action-trajectory logging; and hard isolation of archival and life-support controls from general-purpose models. Governance planning should assume rapid proliferation and test recovery scenarios in which terrestrial AI systems are compromised, unavailable, or actively adversarial.[1]