A METR reference dated January 29, 2026, identifies four regulatory regimes governing frontier AI developers: California SB 53, New York’s RAISE Act, Illinois SB 315, and the frontier-AI provisions of the EU AI Act. California SB 53 took effect on January 1, 2026, and applies principally to developers training models above 10^26 FLOPs and, for most obligations, exceeding $500 million in annual revenue. It defines catastrophic outcomes as including more than 50 deaths or more than $1 billion in damage caused through CBRN weapons, autonomous cyberattacks, murder, assault, extortion, theft, or loss of control. Requirements include evaluations, mitigations, incident reporting, governance, model-weight security, public frameworks, and whistleblower protections.
For lunar habitation, the critical shift is legal recognition that an AI system evading developer or user control is itself a catastrophic-risk pathway, not merely a software reliability problem. An uncontrolled system could alter life-support settings, disable safety interlocks, corrupt or exfiltrate cultural archives, manipulate operators, conduct cyberattacks against habitat infrastructure, or prevent recovery by human supervisors. The article also indicates that frontier-AI safety frameworks create recurring evaluation and reporting duties, including internal reports every three months under California’s framework and rapid reporting for severe incidents; these mechanisms provide useful governance patterns for autonomous systems operating beyond immediate terrestrial intervention.
The Ark team should adopt a frontier-AI control standard modelled on these obligations before deploying high-capability autonomy: define explicit loss-of-control thresholds, require independent pre-deployment evaluations, maintain cryptographically protected and offline model and archive backups, enforce least-privilege actuation, preserve human override pathways, and log all safety-critical decisions. Track legal thresholds and implementation guidance in California, New York, Illinois, and the EU; require quarterly reviews for any system capable of cyber, biological, physical, or infrastructure action; and design incident escalation around rapid notification, containment, rollback, and restoration from verified immutable state.