On 14 September 2026, China’s National Technical Committee 260 on Cybersecurity, under the Cyberspace Administration of China, released version 3.0 of its Artificial Intelligence Safety Governance Framework. The third edition, following the 2024 and 2025 versions, is China’s first national framework with a dedicated agentic-AI risk-management structure. It adds risks involving identity and permission abuse, reasoning and planning failures, tool-call execution, and memory storage, alongside new categories for embodied AI and cybersecurity impacts, including autonomous cyberattack behaviour.[1]
The framework recognizes that AI has shifted from generating answers to autonomously executing tasks. For a lunar settlement, this expands the failure surface from incorrect information to unauthorized command execution, corrupted long-term memory, unsafe planning, compromised robotic systems, and autonomous attacks against life-support, power, communications, manufacturing, or archival systems. Embodied-AI and tool-use controls are therefore survival mechanisms, not merely compliance requirements.[1]
The Ark team should incorporate the framework’s categories into its AI assurance architecture: enforce least-privilege identities, explicit authorization for every tool call, immutable audit logs, segregated memories, bounded planning horizons, human or independently verified approval for irreversible actions, and adversarial tests for autonomous cyberattack behaviour. Monitor subsequent Chinese standards and compare them with international agent-safety regimes, particularly for interoperable controls governing AI agents that operate across settlement, spacecraft, and robotic networks.[1]