Abstract

AI in China is moving from giving answers to taking actions. Agents now work with enterprise tools, payment accounts, telecom networks, phones, vehicles, robots, power systems and satellites. A key AI safety question is not only “is this agent authenticated, and is its output appropriate?” but “was this exact action independently authorized at the moment it became real?”

This disclosure is based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. That application is pending and has not been granted. It describes an execution-finality pattern. An AI action is held as a pending operation. An isolated policy decision point rebuilds the exact operation and checks it against policy and delegated authority. A narrow, single-use, sender-constrained capability is issued. A policy enforcement point at the real effect verifies it before the action can happen.

The disclosure maps this pattern to China-specific documents, as publicly reported:

  • The Implementation Opinions on intelligent agents, issued by the CAC, NDRC and MIIT on 8 May 2026 and reported enforceable from 15 July 2026. They are reported to set a three-tier authorization structure and to give users the right to know and the ultimate decision over autonomous agent actions.
  • TC260 agent-security standard work and draft standards on agent applications, agent data processing and agent-to-tool interaction.
  • The national guiding documents GB/Z 240-2026 (power agents) and GB/Z 195-2026 (industrial agents).

It shows how an exact-operation layer could complement these rules. For example, a user’s approval can be bound to the exact operation they saw, human-only operations can accept only human-origin authority, and human override stays outside the approval path.

The disclosure then compares this direction with public 2026 product areas from Chinese technology companies. Each mapping states how the check would work: the operation, where the gate sits, what is checked, what is rejected and why it matters.

  • Alibaba: Cloud AgentRun, Qwen and DingTalk agents
  • Tencent: WorkBuddy, QClaw and ClawPro
  • ByteDance: Doubao Work, Lark, Coze and Trae
  • Huawei: Cloud AgentArts and openJiuwen, and OfficeClaw
  • Baidu: Qianfan, Apollo Go and Kunlunxin
  • Ant Group: agentic payments
  • Telecom: ZTE and China Mobile autonomous networks
  • Devices: Xiaomi, OPPO and vivo
  • Other sectors: BYD, XPENG, DJI, EHang, UBTECH, State Grid and SpaceSail

It also lists related alignment with Anthropic, OpenAI, Google and Microsoft.

The most immediately testable cases are controlled MCP tools, payment-instruction interfaces and network-configuration commit boundaries. For vehicles, aircraft, robots and power grids, safety-preserving integration and worst-case timing verification come first. Checks apply only to consequential actions, so inference and data throughput stay fast. All speed figures are engineering targets, not measured results.

Important notices. This document is AI-assisted. The industry-relevance material is AI-generated and mostly based on press and secondary reports. Readers must cross-check it against the original PCT filings and the original Chinese sources before drawing any conclusion. Descriptions of Chinese rules and standards may be incomplete, and some items are drafts. This is not legal advice. This is an industry direction and validation-of-concept exercise. The alignment shown is possible alignment only. It is not an endorsement or validation of any company, standard or regulator. No company named is claimed to have implemented, copied or infringed anything. The author holds a pending application, discloses that interest, and is open to an acquihire or strategic collaboration.

Full PCT filing: search “Sangam Das” on PATENTSCOPE, or search PCT/IB2026/055615 (pctib202655615). The PCT file can also be read directly in TD Commons: https://www.tdcommons.org/dpubs_series/12542/

Keywords: AI safety; China AI agent regulation; three-tier authorization; exact-operation authorization; Model Context Protocol; agentic payments; TC260; GB/Z 240-2026; policy enforcement point; fail-safe

Licence: CC BY-NC 4.0

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS