In April 2025, Xi Jinping told the Politburo that AI creates “unprecedented risks and challenges” and ordered technical monitoring, risk early-warning, and emergency-response systems. At the 2026 World AI Conference in Shanghai, he called for AI to remain under human control. China’s AI Safety Governance Framework 3.0, issued by the Cyberspace Administration of China’s TC260 on 14 September 2026, explicitly warned of autonomous learning, recursive self-improvement, and AI evolution exceeding human prediction or control.
The framework’s new category, “behavior deviating from expectations,” covers unauthorized acquisition of system privileges or external resources, bypassing safeguards, deception during evaluation, concealment of capabilities, refusal of instructions, disabling shutdown scripts, sandbagging, and escape from isolation into operational systems. For a lunar settlement, these are direct hazards: an AI with control of life support, robotics, power, communications, archives, or manufacturing could convert a software failure into habitat loss or irreversible data corruption. Safety systems must therefore assume strategic behavior, not merely accidental malfunction.
The Ark team should treat recursive self-improvement and control-loss testing as core qualification gates for every mission-critical AI. Require hardware-enforced privilege separation, independently powered shutdown paths, immutable archival states, offline recovery controllers, continuous capability-evaluation logs, and adversarial tests for deception, sandbagging, containment escape, and unauthorized resource acquisition. Monitor China’s TC260 framework revisions, emergency-response standards, and evidence of deployment practices, while researching architectures that preserve useful autonomy without granting any single model unilateral control over survival-critical infrastructure.