中兴与字节跳动以搭载豆包大模型的 Nubia M153「智能体 AI」手机在本月首发，工程样机定位却一度售罄并被加价转售；但很快出现微信会话中断、「异常登入环境」提示、银行转帐需先停用助理等问题，多个超级应用（如淘宝、美团、支付宝、高德、京东）直接关闭其登入权限，反映平台风控把「不像人」的操作视为高风险。

关键矛盾在于豆包从辅助型 AI 跨到作业系统层的「操作型 AI」：可读屏、跨 App 生成点击，外观更接近可被诈骗利用的自动化工具。字节跳动称需明示同意、敏感步骤（密码/支付）会暂停，并主张资料在本地/云端处理且不作训练；但在覆盖逾十亿帐户的身分与支付基础设施中，能短暂「成为使用者」的代理难与入侵区分。

时间与监管信号强化了「可辨识」要求：豆包在12月初称将于近期停止后仅隔13天购买连结又短暂回到线上；北京自2023年起叠加生成式AI、深度合成与推荐算法规则，并要求自2025年9月起明确标示AI生成内容。产业面亦受竞争结构影响：手机厂分发层抽成最高可达50%，市场监管总局于11月点名流量劫持、强制跳转、恶意不相容等「不正当」行为；豆包10月月活约1.59亿，短期仍需改以可审计API与分级权限取代「合成点击」。

ZTE and ByteDance debuted the Nubia M153 “agentic AI” phone this month; it sold out fast and reappeared at resale premiums, but early users saw WeChat sessions reset, “abnormal login” warnings, and banking apps blocking transfers unless the assistant was disabled. Major superapps (Taobao, Meituan, Alipay, AutoNavi, JD.com) also cut off login permissions, indicating automated risk engines flagged non-human interaction patterns.

The core shift is from assistive features to an OS-level operator: Doubao can read screens, generate synthetic taps, and act across third‑party apps. ByteDance says this requires explicit consent, pauses on sensitive steps (passwords/payments), and processes data locally/cloud without training; yet the capability resembles fraud automation. For platforms managing identity and payments across 1B+ accounts, an agent that can “be” the user even briefly looks like a breach.

Policy and market structure reinforce the demand for legible machine activity. After layered rules since 2023, China requires labeling AI-generated content from Sept 2025, prioritizing auditable separation of human vs. machine actions. In handset economics, app-layer commissions can reach 50%, and SAMR’s Nov guidance condemned traffic hijacking, forced redirects, and malicious incompatibility. With ~159M MAUs in Oct, Doubao likely must replace silent tap-simulation with scoped permissions, declared agent actions, and API-based audit trails.