Reading the 2026 National Cybersecurity Week Agenda as a Compliance List for Foreign Companies

  • The 2026 National Cybersecurity Week ran from 14 to 20 September under the theme of cybersecurity for the people in the intelligent age, with ten central departments as joint sponsors and the opening ceremony in Jinan.
    2026 年国家网络安全宣传周于 9 月 14 日至 20 日举行,主题为「网络安全为人民,网络安全靠人民——智能时代 网安护航」,由十部门联合主办,开幕式在济南举行。
  • The week is neither a compliance obligation nor a new rule; it creates no filing duty and moves no threshold, and treating it as a regulatory event is the first error to avoid.
    宣传周既不是合规义务,也不是新规;它不创设申报义务,也不移动任何门槛,把它当成一次监管事件是首先要避免的错误。
  • Its diagnostic value lies in the agenda: closed-door sessions on critical information infrastructure, open-source supply chain and cloud service security, plus forums on automotive data and personal information.
    它的诊断价值在议程里:关键信息基础设施、开源软件供应链、云计算服务安全三场闭门会,以及汽车数据与个人信息两场分论坛。
  • MLPS Grading is the entry ticket, and it runs as five steps — grade, file, build out, test, supervise — with testing expected from level two and annually from level three.
    等级保护是入场券,按定级、备案、建设整改、测评、监督检查五步推进,二级以上应测评,三级以上原则上每年一次。
  • A Supply Chain Review question arises from two triggers, not from foreign ownership: critical information infrastructure procurement, and a platform with more than a million users listing abroad.
    供应链安全审查的触发点是两个,而不是外资身份:关键信息基础设施采购,以及掌握超过一百万用户个人信息的平台赴国外上市。
  • Data Classification decides which export route is available, and a handler of important data owes an annual risk assessment report to the provincial authority.
    数据分类分级决定可走哪条出境路径,重要数据处理者还须每年向省级主管部门报送风险评估报告。
  • Incident Reporting is where a global playbook and a local clock diverge, because the local duty does not wait for a fully confirmed root cause.
    事件报告是全球预案与本地时限分岔之处,因为本地义务并不等到根本原因完全确认才启动。

Reading the 2026 National Cybersecurity Week Agenda as a Compliance List for Foreign Companies | 从 2026 年国家网络安全宣传周议程读出外资企业的合规清单

What the week is, and what it is not

The 2026 National Cybersecurity Week ran from 14 to 20 September, organised jointly by ten central departments — the Publicity Department, the Cyberspace Administration, the Ministry of Education, the Ministry of Industry and Information Technology, the Ministry of Public Security, the People’s Bank of China, the National Radio and Television Administration, the All-China Federation of Trade Unions, the Communist Youth League Central Committee and the All-China Women’s Federation. The opening ceremony was held in Jinan on 14 September. The theme was cybersecurity for the people, cybersecurity by the people, with the intelligent age and its escort appended as the 2026 suffix.

The first thing to say about the week, and the thing that compliance teams most often get wrong, is that it is a campaign and not an instrument. It creates no filing obligation, imposes no new duty, moves no threshold and repeals nothing. No system has to be registered because the week occurred, and no assessment has to be refreshed because a forum was held. A company that treats the week as a regulatory event will spend September producing documents that no one asked for, and will miss the thing the week is actually good for.

The confusion is understandable, because China’s cybersecurity regime does produce new instruments at a steady rate, and a company that has learned to scan the calendar for new rules will scan this one too. But the regime’s density is exactly why a campaign matters. When the instruments are numerous, the binding constraint on a compliance function is not information about the rules; it is a view of which rules the supervisor currently considers live. A rule that has been in force for four years without a public enforcement story is a different risk from a rule that has been in force for four years and appears twice on a forum agenda. The week is the cheapest available signal of that difference, and it is free.

There is also a practical asymmetry worth noting. A new rule arrives with a text, a commencement date and usually a transition period, which means it can be planned. A shift in supervisory emphasis arrives without any of those things. It shows up as a question in an inspection that the company did not expect, and by then the answer has to be produced on the spot. Reading the agenda is a way of anticipating the question rather than answering it under time pressure.

What the week is good for is diagnosis. The agenda is published, the topics are chosen, and the choice is made by the same departments that write the rules and run the inspections. A programme of fifteen thematic sub-forums, with three topics taken in closed session, is not a compliance requirement, but it is a reasonably reliable statement of what the regulators believe is currently unsolved. Read that way, the week is the one moment in the year when the enforcement agenda is visible in advance and in public.

There is a second, softer function. In practice the week is also the period in which agencies schedule briefings, publish guidance and, in some sectors, run inspection campaigns. Nothing in the campaign documents makes any of that a legal obligation, and a company should not treat the calendar as a deadline. But a company that has a genuine open item — an unfiled system, an overdue test, a data flow with no route — should expect that September is a month in which someone may ask about it.

The distinction matters because it determines what a compliance function should do with the week. It should not build a workstream. It should take the agenda, hold it against its own programme, and use the gaps as a prioritisation input for the year that follows.

Reading the agenda as a map

The 2026 programme is unusually legible. Alongside the opening ceremony and the technology summit, there were fifteen thematic sub-forums, and the titles describe the areas where the organisers think work remains.

Three were closed-door sessions rather than open forums: one for chief security officers of critical information infrastructure operators, one on open-source software supply chain security, and one on cloud service security. Closed-door format is itself information. These are the topics where the discussion is more useful than the announcement, and where the participants are expected to speak about their own arrangements rather than listen to a policy summary.

The open forums covered a wider set: collaborative defence, security insurance, government information system security management, awareness and literacy, personal information protection, artificial intelligence applied to cybersecurity, artificial intelligence security governance, electronic product information erasure, and automotive data security. The opening ceremony released the third version of the artificial intelligence security governance framework, together with the results of an application testing programme for AI-enabled security and a batch of registered security labels for consumer connected cameras.

The three closed-door topics are worth separating out, because each one maps onto a question that a foreign-invested group tends to answer too quickly. Critical information infrastructure is the question of status: whether the entity has been designated, whether it knows, and what follows if it has. Designation is not self-assessed and not published in a list that a company can read; the practical consequence is that a group should be able to say what it would do if the designation arrived, rather than assert that it will not. Open-source supply chain is the question of composition: whether the company can produce, for a system it operates, the list of components it contains and where they came from. Most enterprises cannot, and the reason is not negligence but the fact that modern software is assembled rather than written. Cloud service security is the question of boundary: whether the company knows which of its data sits in a service it does not control, in which region, under whose key management.

Taken together the three closed-door sessions describe a single shift. The supervisory question has moved from whether a company has deployed controls to whether it can describe what it runs. That is a harder question, because it cannot be answered by a policy document.

Two readings follow. The first is that artificial intelligence has moved from principle to instrumentation: a framework in its third version, a testing programme with published results, and a labelling scheme for a consumer device category are the marks of a regime that is being operated rather than declared. The second is that the supply chain has become a first-class subject. Open-source components, cloud services and connected devices are three faces of the same question — whether the thing a company runs is the thing it thinks it runs — and the programme put all three on the agenda in the same week.

For a foreign-invested company the practical move is a mapping exercise, not a reading exercise. Take the eight or nine topics, and for each one ask a single question: what would an inspector ask me for, and could I produce it this week. The answers sort themselves quickly into three groups — items that are documented and current, items that are documented and stale, and items that exist only as an intention. The third group is the work plan.

What follows is that mapping, organised as the five things a foreign-invested enterprise is expected to have, with the single most common point of failure attached to each. The list is not novel. What is useful is the pairing: each obligation has one failure mode that recurs often enough to be worth naming.

Step one and two: the inventory and the grade

The first thing an enterprise is expected to have is an inventory of its systems and data, and the second is a grade for each system that matters. They are treated together because the second is worthless without the first.

An inventory that satisfies an inspector is not an asset register maintained by the finance function and not a service catalogue maintained by the information technology function. It is a list of systems that process data, each with an owner, a hosting location, a network boundary, a data category and a downstream recipient. The reason the distinction matters is that the three lists diverge, and the divergence is exactly where the findings are. A marketing automation tool bought by a regional office, a dealer portal run by a distributor, a plant execution system connected to a group manufacturing platform, a customer service tool embedded in a global licence — none of these necessarily appear in a service catalogue, and all of them process data.

A usable inventory has five fields per system, and the fifth is the one companies omit. The first is the operator, meaning the legal entity that runs the system and would be named in a filing. The second is the hosting location, physical or cloud, including the region. The third is the network boundary, meaning what the system can reach and who can reach it. The fourth is the data category, at the level of granularity the company actually uses. The fifth is the downstream recipient: the group platform, the vendor, the distributor or the overseas service that receives data from the system, and the mechanism by which it does so. Without the fifth field the inventory answers a cybersecurity question and not a data question, and the two regimes described below cannot be joined.

The most common single failure at this step is an inventory built from the organisation chart rather than from the network. The consequence is predictable: the systems that are missed are the ones that were adopted without central approval, and those are also the ones with the weakest controls. A second, related failure is a mismatch between the entity on the registration and the entity that operates the system. A group may file under its China holding company while the operating subsidiary is the actual operator, which produces a filing that is formally complete and substantively wrong.

Grading follows. The level assigned to a system turns on the object that would be harmed by its compromise and the degree of harm, and the discipline of the exercise is that the answer has to be argued rather than assumed. Once a system is graded at level two or above, it is filed with the public security authority, and the filing is the point at which the grade becomes a matter of record.

The most common single failure at this step is grading too low. A manufacturing execution system, an ERP instance holding production and cost data, or a logistics platform serving a large number of counterparties is often graded as an internal system when its compromise would reach beyond the enterprise. The second failure is grading once and never again. A system that was level two when it held internal records may be level three after it is connected to a customer portal, and the change is not always accompanied by a re-grading. The grade is a statement about the current system, not about the system as originally built.

The legal floor matters here. The Regulations on the Administration of Network Data Security require a network data processor to strengthen data security protection on the basis of cybersecurity graded protection. That phrase does the work of connecting two regimes that companies often run as separate programmes: the grade is not only a cybersecurity filing, it is the baseline on which the data security obligations are built. A company that treats graded protection as an information technology chore and data security as a legal chore ends up with two programmes that disagree about which systems exist.

Step three: testing that produces evidence

The third thing an enterprise is expected to have is a test result, and the fourth is evidence that the findings were closed. These are usually described as one step. They are two, and the second is the one that is examined.

Testing is a measurement at a point in time, carried out by an institution with the relevant qualification, against a defined set of control requirements. The output is a report with a score and a list of findings, sorted by severity. The report has a limited life, and a company that is asked for it in September and produces one from two years ago has produced evidence of a problem rather than evidence of compliance.

Remediation is the part that is examined, and it is examined through the file rather than through the system. The question an inspector asks is not whether a finding was fixed but whether the company can show that it was fixed: a change record, a configuration screenshot with a date, a re-test result, a compensating control with a documented rationale and a review date. A finding closed by assertion is a finding that will be reopened.

What an inspector samples is worth describing, because it shapes what the file should contain. In practice the sample is small and deep: two or three findings, followed from the report through the change record to the re-test. A company with a thin file on three findings looks worse than a company with a thick file on ten, because the inspection is testing whether the process works rather than whether the score is high. The implication is counter-intuitive and worth stating: it is better to have a small number of fully evidenced closures and an honest list of open items than a clean report with no trail behind it.

The most common single failure at this step is a remediation file that contains a plan but no closure. The second is a gap analysis that was never costed, so the high-severity items were deferred year after year and the test report became a document that everyone signs and no one acts on. The third, less obvious, is that the remediation is scoped to the tested system rather than to the class of systems. A finding about an unpatched component on one server is, almost always, a finding about the patch process, and fixing the server leaves the process untouched.

There is a sequencing point worth making because it is often reversed in practice. Build-out comes before testing in the standard order of the five steps, and the order is not decorative. A test run before remediation measures the old design and produces findings that will be found again next year. Companies that test early because a procurement deadline requires a report, and then remediate afterwards, end up paying for two tests and one report that describes a system that no longer exists.

Step four: the Supply Chain Review question

The fourth thing an enterprise has to be able to answer is whether a procurement or a listing triggers a cybersecurity review. This is the item most often answered by assertion rather than by analysis.

The review measures of 2022 set out two triggers. The first is procurement of network products and services by an operator of critical information infrastructure, where the procurement affects or may affect national security; the operator is required to make that prediction itself and to file where the answer is affirmative. The second is a network platform operator holding personal information of more than one million users that seeks to list abroad, which must file. The definition of network products and services is broad and specific at the same time: core network equipment, important communications products, high-performance computers and servers, large-capacity storage, large databases and application software, cybersecurity equipment, cloud computing services, and other items with a significant bearing on critical information infrastructure, cybersecurity and data security.

Two features of the regime are worth stating plainly. The review is not a prohibition, and it is not a general licence for procurement. It is a process with a defined sequence and defined clocks: a decision on whether review is needed, an initial review with a stated period, a special review procedure for cases where the member departments disagree, and a set of risk factors that the review is required to assess. And the review is not the only screening that may apply: the measures provide expressly that where data security review or foreign investment security review is separately regulated, those requirements apply as well.

The uncertainty that companies find hardest to manage is designation. Critical information infrastructure is identified by the protection departments of the State Council under the regulations, and the identification is communicated to the operator rather than published as a register. A foreign-invested company therefore often cannot prove a negative, and the honest position is not “we are not designated” but “we have not been notified, and here is what we would do if we were”. That position is defensible, and it is more useful than an assertion, because it forces the company to answer the follow-up question in advance: who would run the review, what would the procurement inventory look like, and how long would a filing take.

The most common single failure at this step is to answer the trigger question with the wrong fact. Foreign ownership is not a trigger; it is a reason to expect more questions, but the legal test is the operator status and the procurement. A company that concludes it is outside the regime because it is foreign-invested has answered a question that was not asked. The second failure is to treat the review as a vendor due diligence exercise and to stop there. The measures require the operator to require, through procurement documents and contracts, that the provider commit not to use its position to obtain data unlawfully, not to control or manipulate the customer’s equipment without cause, and not to interrupt supply or necessary technical support without justification. Those three commitments are contract terms, they cost nothing to insert, and they are frequently absent from the template agreements that a group uses globally.

A third failure is more subtle and worth flagging for groups that procure centrally. Where the contracting entity is outside China and the operating entity inside China uses the service, the Chinese operator may not see the procurement at all. The trigger test attaches to the operator’s use of the product, not to who signed the purchase order.

Step five: Data Classification and the export route

The fifth thing an enterprise needs is a classification of its data, because every export route depends on it.

The statutory base is the Data Security Law, which establishes a system of classified and graded data protection and a catalogue of important data. The 2025 network data regulation fills in the operating detail: the national coordination mechanism coordinates the making of the important data catalogue, regions and departments determine their own catalogues, processors identify and declare important data, and where data is confirmed as important the responsible region or department must notify the processor or publish the determination. That last provision matters in both directions. A processor is not required to guess, but a processor that has been told, or that operates in a sector where the catalogue has been published, cannot claim ignorance.

The regulation adds two obligations that a foreign-invested group should read together. A handler of important data must carry out an annual risk assessment of its network data processing activities and submit a report to the provincial-level competent authority, and the report must cover, among other things, the data export situation. And a processor handling personal information of ten million individuals or more is made subject to the provisions that apply to handlers of important data on security officers, management bodies and reporting. The second provision catches consumer-facing businesses that would not describe themselves as handling important data at all.

The export route then follows from the classification. Personal information may leave under a security assessment, a certification, a standard contract, or one of the statutory exemptions; important data must go through a security assessment. The free trade zone negative lists discussed elsewhere in this collection change the routing inside a zone, not the classification.

The classification exercise itself is usually run on three axes, and the order matters. The first axis is the object: whether the data concerns national security, economic operation, public health, personal rights or commercial interest. The second is the degree of harm, which is where most arguments occur, because harm is easy to assert and hard to bound. The third is the scale, which for personal information is a count of individuals and for other data is a matter of coverage and precision. A classification that starts from the third axis — counting records — will produce a defensible answer for personal information and a poor one for everything else, which is the shape of most classification failures in industrial and trading companies.

The most common single failure at this step is to equate personal information with data. A company that has correctly concluded that a flow carries no personal information may still be exporting important data, because important data is defined by its significance rather than by whether an individual is identifiable. Production yields, capacity utilisation, inventory levels, pricing positions and supply chain maps are the standard examples, and they travel in the same integration feeds as the innocuous data around them. The second failure is to treat a transfer within the group as not a transfer. The test is whether the data leaves the territory, not whether it leaves the legal entity.

Incident Reporting, and a calendar for the year

The obligation that ties the other five together is reporting, because it is the moment at which everything else becomes visible to someone outside the company.

The duties sit in several instruments and they are cumulative. The Cybersecurity Law, as amended with effect from 1 January 2026, provides the general framework and the penalties; the amendment raised the penalties for a range of network security violations and extended the law’s reach. The network data regulation adds specifics: where a defect or vulnerability in a product or service endangers national security or the public interest, the processor must report to the competent authority within twenty-four hours; where a data security incident occurs, the processor must activate its contingency plan, take measures to prevent the harm from spreading and report as required, and where the incident harms the rights of individuals or organisations it must notify the affected parties. The regulation also requires a processor to report suspected criminal conduct to the public security and state security authorities.

The most common single failure at this step is a collision of clocks. A group that operates under a seventy-two-hour notification regime, and under an internal rule that notification follows confirmation of a root cause, will find that the local duty is not built that way: the trigger is the occurrence of the incident, the report is due as required by the authority’s rules, and waiting for confirmation is a decision to report late. The second failure is authority. In many groups the local entity cannot report without clearance from a regional or global function, and the clearance process takes longer than the local window. The fix is a pre-agreed protocol that names who may report, on what trigger, without further approval.

Such a protocol has four elements and takes about a day to write. It names the reporting entity and a deputy. It defines the trigger in local terms rather than in the group’s terms, so that an incident is reportable when it meets the local threshold and not only when the group’s severity scale says so. It states that the local report may go out before the group’s own assessment is complete, and that the group will be informed in parallel rather than in advance. And it records who is authorised to speak to the authority, because a report is a statement by the company and a second, inconsistent statement is worse than a late one.

Set against those five items, the year has a shape. Graded protection testing recurs annually for level three systems and on the cycle applicable to level two. The important data risk assessment is annual. The personal information compliance audit is periodic. Filings have to be updated when a system changes grade or a new system is brought into scope. Contingency plans have to be exercised, not merely written. A company that lays these on a calendar discovers that the workload is uneven and that September, the month of the campaign, is a poor month to add anything.

Which returns the argument to where it began. The week does not create the obligations and does not change them. It publishes, once a year, the list of things the supervisors are still worried about, and it does so in a form that a compliance function can use as a check rather than as a task. The value is in the check. A company that reads the agenda, holds its own programme against it, and finds that four of the five items are documented and current has learned something worth knowing. A company that reads it as a deadline has learned nothing and will spend the month producing paper.

Sources

  • 国家网络安全宣传周官网·2026 年国家网络安全宣传周开幕式在山东济南举行(2026-09-14):https://www.gjwlaqxcz.cn/web/news/detail.html?id=235
  • 中国网信网·中华人民共和国主席令(第六十一号)——《全国人民代表大会常务委员会关于修改〈中华人民共和国网络安全法〉的决定》自 2026-01-01 起施行(2025-10-28):https://www.cac.gov.cn/2025-10/29/c_1763461514706581.htm
  • 新华网·引导人工智能发展——网络安全法完成修改(2025-10-28,含罚则调整要点):https://www.news.cn/politics/20251028/b3400522666a4b3f8c844b9fa6c31131/c.html
  • 中国网信网·《网络安全审查办法》(国家互联网信息办公室等十三部门令第 8 号,2022-02-15 施行):https://www.cac.gov.cn/2022-01/04/c_1642894602182845.htm
  • 中国政府网·《网络数据安全管理条例》(国务院令第 790 号,2025-01-01 施行):https://www.gov.cn/gongbao/2024/issue_11646/202410/content_6980863.html

从 2026 年国家网络安全宣传周议程读出外资企业的合规清单

宣传周是什么,不是什么

2026 年国家网络安全宣传周于 9 月 14 日至 20 日举行,由中央宣传部、中央网信办、教育部、工业和信息化部、公安部、中国人民银行、国家广播电视总局、全国总工会、共青团中央、全国妇联十部门联合主办。开幕式于 9 月 14 日在济南举行。主题是「网络安全为人民,网络安全靠人民——智能时代 网安护航」。

关于这一周,第一句要说清、也是合规团队最常搞错的一句是:它是一场宣传活动,不是一件监管工具。它不创设申报义务,不新增任何职责,不移动任何门槛,也不废止任何规定。没有哪个系统因为宣传周举行而需要登记,也没有哪份评估因为开了论坛而需要重做。把宣传周当成一次监管事件的企业,会在九月生产一批没人要的文件,同时错过这一周真正有用的地方。

这种混淆可以理解,因为中国的网络安全制度确实在以稳定的节奏出文件,而一家已经养成「扫日历找新规」习惯的企业会把这个日历也扫一遍。但正因为制度密集,宣传活动才有意义。当规范多到一定程度,合规职能的约束条件就不再是「规则的信息」,而是「监管方当下认为哪些规则是活的」这一判断。一条施行四年、却从未有过公开执法故事的规定,与一条施行四年、在议程上出现两次的规定,风险不一样。宣传周是能拿到这个差别的最便宜信号,而且免费。

还有一处实际的错位值得注意。新规到来时带着文本、施行日,通常还有过渡期,这意味着它可以被规划。监管重心的移动不带这些东西中的任何一样。它表现为检查中一个企业没预料到的问题,而到那时答案必须当场给出。读议程,是提前想到那个问题,而不是在压力下回答它。

它真正有用的地方是诊断。议程是公开的,议题是选出来的,而选议题的人就是写规则、跑检查的那几个部门。十五场专题分论坛、其中三个议题以闭门会形式进行,这不是合规要求,但它是一份相当可靠的陈述:监管方认为哪些问题目前尚未解决。这样读,宣传周就是一年之中唯一一次,执法议程提前且公开可见的时刻。

还有一层更软的功能。实践中,宣传周也是各部门集中安排宣讲、发布指引,并在某些行业开展检查的时间段。宣传周的文件里没有任何一条把这变成法律义务,企业也不该把日历当成截止日。但一家手里确实有未了事项的企业——一个没备案的系统、一次逾期的测评、一条没有路径的数据流——应当预期九月是可能有人来问的月份。

这个区分之所以要紧,是因为它决定合规职能该拿这一周做什么。它不该为此立项。它该把议程拿过来,对一遍自己的方案,把差距当作下一年度的排序输入。

把议程当地图读

2026 年的议程异常清晰。除开幕式与技术高峰论坛之外,共有十五场专题分论坛,标题本身就说明了主办方认为哪些地方还有活要干。

其中三场是闭门会而不是公开论坛:一场面向关键信息基础设施运营者的首席安全官,一场谈开源软件供应链安全,一场谈云计算服务安全。闭门这个形式本身就是信息。这些议题上,讨论比宣布更有用,与会者被期待讲自己的安排,而不是听政策综述。

公开论坛覆盖面更宽:协同防御、网络安全保险、政务信息系统安全管理与实践、意识和素养提升、个人信息保护、人工智能技术赋能网络安全应用、人工智能安全治理、电子产品信息清除、汽车数据安全。开幕式发布了《人工智能安全治理框架》第三版,同时发布了人工智能技术赋能网络安全应用的测试结果,以及一批消费类网联摄像头的网络安全标识备案产品。

三场闭门会的议题值得单拎出来,因为每一个都对应一个外资集团往往回答得太快的问题。关键信息基础设施是「身份」问题:主体有没有被认定、知不知道、被认定了之后要做什么。认定不是自我评估,也不在一份企业可以查的名单上;实际的推论是,集团应当能说清「认定真的来了我会怎么做」,而不是断言它不会来。开源软件供应链是「构成」问题:对一个自己在运行的系统,公司能不能列出它包含哪些组件、这些组件从哪里来。多数企业列不出来,而原因不是疏忽,是现代软件是装配出来的、不是写出来的。云计算服务安全是「边界」问题:公司知不知道自己的哪些数据放在一个它不控制的服务里、在哪个区域、由谁做密钥管理。

三场闭门会合起来描述的是一次位移。监管的问题,从「企业有没有部署控制措施」移到了「企业能不能描述自己在运行什么」。后一个问题更难,因为它无法用一份制度文件回答。

由此可以读出两件事。第一,人工智能已经从原则走到了工具层面:一个出到第三版的框架、一个有公开结果的测试项目、一套针对某一消费设备类别的标识备案,这些都是一套制度在被运行、而不只是被宣告的标志。第二,供应链成了一等议题。开源组件、云服务、联网设备是同一个问题的三张面孔——企业运行的东西,是不是它以为的那个东西——而议程把这三张面孔放在同一周里。

对外资企业而言,实际动作是一次映射,而不是一次阅读。把八九个议题拿过来,对每一个只问一个问题:检查的人会向我要什么,我这周拿得出来吗。答案很快会分成三堆:有文件且是最新的、有文件但已过时的、只停留在打算上的。第三堆就是工作计划。

下面就是这次映射,按外资企业被期待具备的五件事组织,每一件配一项最常见的失分。这份清单并不新颖。有用的是配对:每一项义务都配一个反复出现到值得命名的失败方式。

第一步与第二步:清单与定级

企业被期待具备的第一件东西是系统与数据的清单,第二件是每个重要系统的等级。两者放在一起讲,因为离开第一件,第二件没有意义。

一份经得起检查的清单,不是财务口维护的资产台账,也不是信息技术口维护的服务目录。它是一份处理数据的系统清单,每个系统带负责人、托管位置、网络边界、数据类别与下游接收方。

一份能用的清单,每个系统有五个字段,而第五个是企业最常漏掉的。第一是运营者,即运行该系统、并会在备案中被点名的法律主体。第二是托管位置,物理机房或云端,含区域。第三是网络边界,即该系统能触达什么、谁能触达它。第四是数据类别,颗粒度按企业实际使用的来。第五是下游接收方:接收该系统数据的集团平台、供应商、经销商或境外服务,以及接收所经由的机制。缺了第五个字段,清单回答的是网络安全问题,而不是数据问题,下面要讲的两套制度也就接不上。

这个区分之所以要紧,是因为三份清单会分叉,而分叉处正是问题所在。区域办公室自行采购的营销自动化工具、经销商运营的门户、连着集团制造平台的工厂执行系统、嵌在全球授权里的客服工具——这些都不一定出现在服务目录里,而它们全都在处理数据。

这一步最常见的单项失分,是按组织架构而不是按网络来建清单。后果可以预料:被漏掉的系统,正是那些未经集中审批就被采用的系统,而它们也是控制最弱的系统。第二种相关的失分,是备案主体与系统运营主体不一致。集团可能以中国控股公司名义备案,而实际运营者是业务子公司,于是得到一份形式上完整、实质上错误的备案。

定级随之而来。一个系统定在哪一级,取决于它被侵害时会损害什么客体、损害到什么程度,而这项功课的纪律在于:答案要论证,不能假定。系统定在二级及以上,要向公安机关备案,而备案正是等级成为记录的那一刻。

这一步最常见的单项失分,是定级偏低。制造执行系统、装着生产与成本数据的核心业务系统、服务大量交易对手的物流平台,常被当作内部系统定级,而它们一旦出事,影响会超出企业自身。第二种失分是定一次、此后不再动。一个装着内部记录时是二级的系统,在接上客户门户之后可能应当是三级,而这个变化并不总伴随着重新定级。等级是关于当前系统的陈述,不是关于系统最初建成时样子的陈述。

这里有一条法律底线值得点出。《网络数据安全管理条例》要求网络数据处理者在网络安全等级保护的基础上加强网络数据安全防护。这句话的作用,是把企业常常分开运行的两套制度接上:等级不只是一份网络安全备案,它是数据安全义务赖以搭建的地基。把等级保护当成信息技术的杂活、把数据安全当成法务的杂活的企业,最后会有两套对「哪些系统存在」意见不一致的方案。

第三步:能产出证据的测评

企业被期待具备的第三件东西是一份测评结果,第四件是问题项已关闭的证据。这两件常被说成一步。它们是两步,而被查的是第二步。

测评是一次时点测量,由具备相应资质的机构,对照一套确定的控制要求完成。产出是一份带分数和问题清单的报告,问题按严重程度排序。报告有寿命,一家在九月被要求提供、却拿出一份两年前报告的企业,提供的是问题的证据,而不是合规的证据。

整改才是被查的部分,而查的方式是查卷宗,不是查系统。检查者问的不是问题有没有修好,而是企业能不能证明它修好了:一条变更记录、一张带日期的配置截图、一次复测结果、一项写明了理由与复核日期的补偿性控制。以断言方式关闭的问题项,是会被重新打开的问题项。

检查取样什么,值得描述,因为它决定了卷宗里该放什么。实践中取样量小、深度大:两三个问题项,从报告顺着变更记录追到复测。三个问题项上卷宗很薄的企业,看起来比十个问题项上卷宗很厚的企业更糟,因为检查考的是流程是否运转,不是分数是否漂亮。由此有一个反直觉、但值得说清的推论:少数几项证据完整地关闭、加上一份诚实的未关闭清单,比一份干净却没有痕迹的报告更好。

这一步最常见的单项失分,是有整改计划而没有关闭。第二种是差距分析从未被计价,于是高危项年复一年地被推后,测评报告变成一份人人签字、无人执行的文档。第三种不那么显眼:整改的范围被限定在被测系统,而不是那一类系统。一台服务器上未打补丁的组件,几乎总是一个关于补丁流程的问题,修好那台服务器,流程原封不动。

有一个次序问题值得说,因为实践中常被倒过来。按五步的标准顺序,建设整改在测评之前,而这个顺序不是装饰。在整改之前做的测评,测的是旧设计,产出的问题明年还会再被发现一次。因采购节点需要一份报告而提前测评、事后再整改的企业,最后付两次测评的钱,其中一份报告描述的是一个已经不存在的系统。

第四步:供应链审查这道题

企业必须能回答的第四件事,是某次采购或某次上市是否触发网络安全审查。这一项最常被用断言回答,而不是用分析回答。

2022 年的审查办法给出两个触发点。第一是关键信息基础设施运营者采购网络产品和服务,且该采购影响或者可能影响国家安全;运营者须自行预判,答案为肯定时申报。第二是掌握超过一百万用户个人信息的网络平台运营者赴国外上市,必须申报。网络产品和服务的定义既宽又具体:核心网络设备、重要通信产品、高性能计算机和服务器、大容量存储设备、大型数据库和应用软件、网络安全设备、云计算服务,以及其他对关键信息基础设施安全、网络安全和数据安全有重要影响的品类。

这套制度有两个特征值得直说。审查不是禁止,也不是采购的通行证。它是一个有既定顺序和既定时钟的流程:先确定是否需要审查,再进入有明确期限的初步审查,成员单位意见不一致时进入特别审查程序,并配有一组必须评估的风险因素。审查也不是唯一一道程序:办法明确规定,数据安全审查、外商投资安全审查另有规定的,应当同时符合其规定。

企业最难处理的不确定性是认定本身。关键信息基础设施由国务院保护工作部门依规认定,认定结果告知运营者,而不以名录形式公布。因此外资企业往往无法证明一个否定命题,而诚实的立场不是「我们没被认定」,而是「我们尚未收到告知,若收到,我们的动作是这样的」。这个立场站得住,也比一句断言更有用,因为它逼着企业提前回答后续问题:谁来牵头审查、采购清单长什么样、申报要走多久。

这一步最常见的单项失分,是用错误的事实回答触发问题。外资身份不是触发点;它是预期会被多问几句的理由,但法律上的检验是运营者身份与采购行为。一家因为自己是外资就断定不在制度范围内的企业,回答的是一个没有被问的问题。第二种失分是把审查当成供应商尽职调查,做到那里就停。办法要求运营者通过采购文件、协议,要求产品和服务提供者承诺:不利用提供产品和服务的便利条件非法获取用户数据,无正当理由不非法控制和操纵用户设备,无正当理由不中断产品供应或必要的技术支持服务。这三项承诺是合同条款,写进去不花钱,而集团全球模板协议里常常没有。

第三种失分更微妙,值得提示给集中采购的集团。当签约主体在中国境外、而使用服务的主体在中国境内时,中国的运营者可能根本看不到这次采购。触发检验附着于运营者对产品的使用,不附着于谁签了采购单。

第五步:数据分类分级与出境路径

企业需要的第五件东西是数据分类分级,因为每一条出境路径都取决于它。

法律基础是《数据安全法》,它确立了数据分类分级保护制度与重要数据目录制度。2025 年施行的《网络数据安全管理条例》把操作细节补齐:国家数据安全工作协调机制统筹协调有关部门制定重要数据目录,各地区、各部门确定本地区、本部门以及相关行业、领域的重要数据具体目录,数据处理者按规定识别、申报重要数据,确认为重要数据的,相关地区、部门应当及时告知数据处理者或者公开发布。最后这一句在两个方向上都有意义。企业不需要猜,但一家已经被告知、或在目录已发布的行业里经营的企业,不能主张不知情。

条例还加了两项义务,外资集团应当连起来读。重要数据的处理者应当每年度对网络数据处理活动开展风险评估,并向省级以上有关主管部门报送风险评估报告,报告内容中明确包含网络数据出境情况。处理一千万人以上个人信息的网络数据处理者,还要遵守条例对重要数据处理者在安全负责人、管理机构与报告方面的规定。后一条会圈住那些自认为完全不碰重要数据的消费类业务。

出境路径随后由分类分级推出。个人信息可以走安全评估、认证、标准合同或法定豁免之一;重要数据必须走安全评估。其他地方讨论的自贸试验区负面清单,改变的是区内走哪条路,不是分类分级本身。

分类分级这项功课通常按三个轴做,而顺序有讲究。第一轴是客体:数据涉及国家安全、经济运行、公共卫生、个人权益,还是商业利益。第二轴是危害程度,争论多半发生在这里,因为危害容易主张、难以界定。第三轴是规模,对个人信息是个自然人计数,对其他数据是覆盖面与精度。从第三轴起步——先数记录条数——的分类,会给个人信息一个站得住的答案,给其他一切一个糟糕的答案,而这正是工业与贸易企业里多数分类失败的样子。

这一步最常见的单项失分,是把个人信息等同于数据。一家正确判断出某条数据流不含个人信息的企业,仍可能在出境重要数据,因为重要数据的定义看的是重要性,而不是能否识别到某个自然人。产量、产能利用率、库存水平、报价头寸、供应链图谱是标准例子,而它们和周围无害的数据坐在同一条集成管道里。第二种失分是把集团内部传输当成不是传输。检验的是数据是否出境,不是是否出门。

事件报告,与一年的日历

把其余五项串起来的那项义务是报告,因为它是其他一切被公司外部看见的那一刻。

相关义务分散在几部规范里,而且是累加的。《网络安全法》经修改后自 2026 年 1 月 1 日起施行,提供总体框架与罚则;此次修改提高了一系列网络安全违法行为的处罚,并扩大了法律的域外适用情形。条例补上具体:网络产品、服务存在安全缺陷、漏洞等风险,涉及危害国家安全、公共利益的,应当在二十四小时内向有关主管部门报告;发生网络数据安全事件,应当立即启动预案、采取措施防止危害扩大并按规定报告,对个人、组织合法权益造成危害的,还应当通知利害关系人。条例还要求,处置过程中发现涉嫌违法犯罪线索的,向公安机关、国家安全机关报案。

这一步最常见的单项失分,是时钟相撞。一家同时运行七十二小时通报制度、并在内部规定「确认根本原因后再通报」的集团,会发现本地义务不是这么设计的:触发点是事件发生,报告按主管部门规定的时限报送,等待确认等于决定迟报。第二种失分是权限。很多集团里,本地实体无权在未获区域或全球职能放行前报告,而放行流程比本地窗口更长。解法是一份事先约定的规程:写明谁可以报告、在什么触发点报告、无需再行审批。

这样一份规程有四个要素,写下来大约一天。它点名报告主体与一名替补。它用本地口径而不是集团口径定义触发点,使事件在满足本地门槛时就应当报告,而不是只在集团的严重度分级说应当时才报告。它写明本地报告可以在集团自身评估完成之前发出,集团同步获知而不是事先批准。它还记录谁有权对主管部门陈述,因为报告是公司的陈述,第二份不一致的陈述比一份迟到的陈述更糟。

把五项摆在一起,一年就有了形状。三级系统的等级测评按年进行,二级按适用周期;重要数据风险评估按年;个人信息合规审计按周期;系统定级变化或新系统纳入范围时,备案要更新;应急预案要演练,而不只是写出来。把这些排在日历上的企业会发现工作量并不均匀,而九月——宣传周所在的月份——并不是适合再加事情的时候。

这就把论证带回起点。宣传周不创设义务,也不改变义务。它一年一次地公布监管方仍在担心的事情,形式是一份合规职能可以拿来当检查表、而不是当任务单的东西。价值在检查。一家读了议程、把自己的方案对上去、发现五项里有四项有文件且是最新的企业,知道了值得知道的事。一家把它读成截止日的企业,什么也没学到,还会花掉一个月生产纸张。

来源

  • 国家网络安全宣传周官网·2026 年国家网络安全宣传周开幕式在山东济南举行(2026-09-14):https://www.gjwlaqxcz.cn/web/news/detail.html?id=235
  • 中国网信网·中华人民共和国主席令(第六十一号)——《全国人民代表大会常务委员会关于修改〈中华人民共和国网络安全法〉的决定》自 2026-01-01 起施行(2025-10-28):https://www.cac.gov.cn/2025-10/29/c_1763461514706581.htm
  • 新华网·引导人工智能发展——网络安全法完成修改(2025-10-28,含罚则调整要点):https://www.news.cn/politics/20251028/b3400522666a4b3f8c844b9fa6c31131/c.html
  • 中国网信网·《网络安全审查办法》(国家互联网信息办公室等十三部门令第 8 号,2022-02-15 施行):https://www.cac.gov.cn/2022-01/04/c_1642894602182845.htm
  • 中国政府网·《网络数据安全管理条例》(国务院令第 790 号,2025-01-01 施行):https://www.gov.cn/gongbao/2024/issue_11646/202410/content_6980863.html