MIIT Notice No. 42: Every AI Agent Device Must Carry an Identity Code to Access the Network as the World's First Hardware-Layer Regulatory Framework Takes Effect

On November 6, 2026, China's Ministry of Industry and Information Technology issued the "Notice on Strengthening the Network Access Management of Intelligent Agent Terminal Devices" (MIIT General Office Information Management [2026] No. 42). Its core provisions specify that terminals such as smartphones, tablets, and wearable devices pre-installed with AI agents must complete agent adaptation verification, fraud-related risk security assessment, and pre-emptive anti-fraud capability evaluation before obtaining a network access license, and that AI agents used on terminals should be equipped with a unique identity code with reference to the series of national standards for agent interconnection. This is the world's first regulatory framework that mandatorily embeds AI agent identity management into the hardware access approval process for devices.

From "software approval" to "hardware checkpoint" — where this gate is installed

To understand the technical implications of this notice, one must first understand China's "network access license" system. Before any smartphone or tablet can be sold in the Chinese market, it must pass MIIT's network access testing and obtain a license — akin to an "admission permit" at the device level. In the past, this checkpoint mainly reviewed hardware indicators such as radio frequency compliance, baseband chip parameters, and radiation safety standards, and had nothing to do with the software functions running on the device.

What Notice No. 42 does is physically "weld" AI agent compliance checks into this checkpoint that previously looked only at hardware. To obtain a network access license, manufacturers must now not only pass RF testing but also demonstrate that the pre-installed AI agents have completed adaptation verification with the device and passed anti-fraud assessment. In other words, the compliance status of AI agents becomes a precondition for hardware admission; the two are bundled for approval and cannot be separated.

According to Cailianshe, the unique identity code requirement references the "series of national standards for agent interconnection." This set of standards is itself still under development, but its logic resembles the IMEI code of a device — every AI agent instance running on a terminal will have a traceable identity identifier, enabling regulators to trace back to the specific agent version and vendor when problems arise.

A compliance chain made up of four checkpoints

From the full text of the notice, what MIIT has built is not a single inspection node but a four-tier institutional system covering the entire lifecycle of devices.

The first tier is adaptation verification: before network access, AI agents must complete compatibility testing with terminal devices, covering dimensions such as permission invocation, communication capabilities, network security, and data security. This checkpoint requires manufacturers to systematically map out the permission boundaries of AI agents before official market launch — which sensors they can invoke and which system interfaces they can access must all be clearly recorded in the testing documents.

The second tier is network access testing: implemented in tandem with hardware admission approval, it includes fraud-related risk security assessment and pre-emptive anti-fraud capability evaluation. The notice specifically emphasizes that "terminal production enterprises shall ensure the consistency of intelligent agent terminals before and after obtaining the network access license," meaning that the core behavior of AI agents cannot be quietly changed after the license is obtained.

The third tier is live network upgrade management: when pushing upgrades to sold devices via OTA or other means, manufacturers must submit the upgrade plan to MIIT in advance, explaining the upgrade timing, scale, user scope, business scenarios, security measures, and anti-fraud measures; only after expert evaluation and the solicitation of opinions can services be provided to users. This effectively closes off the usual practice of "launch first, fix compliance later."

The fourth tier is distribution management: application developers must explicitly declare whether they agree to be invoked by AI agents; distribution platforms may provide agent installation services only within the scope authorized by developers, and the installation scope must be limited to compatible device models.

Whose costs rise, whose barriers get higher

For terminal manufacturers, the rise in compliance costs is direct. According to 21st Century Business Herald, adaptation verification and network access testing will increase "R&D and mass production costs," and the pre-approval requirement for OTA upgrades means product iteration rhythms must leave buffer time for regulatory processes. For domestic phone manufacturers that have long relied on rapid iteration to seize the market, this is a substantive constraint on pace.

For overseas brands, the challenge is more complex. Manufacturers such as Apple and Samsung that sell devices with pre-installed AI features in the Chinese market must also be incorporated into this framework. This means their globally unified AI agent versions will need to go through separate adaptation verification and fraud assessment processes for the Chinese market, effectively raising the operational threshold for maintaining AI features in China.

For application developers, the mandatory declaration system for "whether to agree to be invoked by AI agents" changes the existing permission ecosystem. In the past, AI agents could access third-party applications by simulating user operations; now distribution platforms must premise this on developers' written authorization. This clause is directly related to the commercial interests of certain super-apps — according to Techie Ray's analysis, it gives applications a legal basis to refuse access to other companies' AI agents.

From another perspective, for large platform-type enterprises that have already invested heavily in compliance, this notice actually builds a threshold that is unfriendly to smaller players: compliance costs are relatively fixed, and the smaller the market share, the higher the per-unit cost allocation.

Two different "passport" philosophies

Notice No. 42 is not an isolated event. Just two days before the notice was issued, on November 4, 2026, China's Cyberspace Administration, together with the global mobile industry association, released an industry self-discipline standard for AI agents at the Wuzhen Internet Conference, forming a dual-track collaborative structure of "soft industry standards + hard device admission."

Almost at the same time, also on November 4, according to Techie Ray, Singapore's Minister of State Jasmin Lau said in Parliament that the Singapore government is jointly testing a completely different technical route with GoDaddy, Infoblox, Cloudflare, and Google: exploring whether the Internet's Domain Name System (DNS) can be used to identify and verify AI agents across platforms and organizations. Under this approach, the "ID card" of an AI agent would not be issued by a national agency but carried by the Internet's existing distributed infrastructure, with the government deciding after evaluation whether an additional "national registration system" is still needed.

These two routes reflect fundamentally different philosophies of technical governance. China's approach anchors trust in the approval actions of a national agency — pass the network access testing, and you hold a "passport" issued by the state; the approach Singapore is testing disperses trust into Internet infrastructure such as DNS, closer to the decentralized tradition of the Western Internet.

The EU's AI Act provides a third comparative reference. The EU's regulatory logic mainly rests on algorithmic transparency and tiered risk assessment, focusing on whether AI systems' decision-making processes are explainable and whether they discriminate against specific groups — this is a governance framework at the content and application layers. China's Notice No. 42 moves the control point forward to the device layer, screening at the hardware admission stage before AI agents even start running. The difference between the two is not one of strictness but a fundamental difference in the access point for control.

Outlook: the early game of the standards contest

When the specific content of the series of national standards for agent interconnection will be released, and how its technical specifications will determine whether the identity code system can be practically implemented and form a traceable regulatory loop; whether Singapore's DNS verification experiment can form a scalable international solution, and how it will coordinate or compete with China's device-layer identity framework. This standards contest over "who issues passports to AI agents" is still in its opening phase.