China Plans AI Safety Standards for the Full Model Lifecycle: What the NDRC Said

China says it will write safety standards for artificial intelligence that cover a model from its data to its deployment. On November 3, 2026, Wang Ruomeng, a senior official at the National Development and Reform Commission (NDRC), told a Beijing briefing that the state planner will speed up legislation on how AI is developed and applied, according to Bloomberg's report syndicated by The Star.

This is a statement of direction, not a published rulebook. No final text, effective date or penalty schedule came with it. Below we separate what was said from what is only implied.

Agents that act on your behalf need rules. See how explainx helps secure your agents →

TL;DR: what was said and what was not

What Wang Ruomeng said

According to Bloomberg, Wang described the approach this way: "We will accelerate legislation to support core AI enablers – data, computing power and algorithms – and regulate key application scenarios, while establishing safety standards and ethical guidelines covering the full life cycle of AI."

Developers and operators, Wang said, should bolster risk management and disclose safety risks, incidents and vulnerabilities.

Bloomberg ties the move to a "spate of recent hacks in which AI went rogue or radically enhanced cyber-breach attempts" and to warnings from Silicon Valley leaders about existential risk. That framing is Bloomberg's context; the article gives no specifics on those hacks. Wang also said, per Bloomberg, that "AI safety is not a problem one company, industry or country can tackle alone."

Regulatory compliance stamp on a document, representing China AI safety standards for the model lifecycle

The balance Beijing is trying to hold

Two things pull in opposite directions. China wants its labs to keep pace with the best models from OpenAI and Anthropic, and it wants to avoid a loss of control or a wave of AI-enabled cyberattacks.

The plan reflects both:

• Constraint: lifecycle standards, risk management and incident disclosure.
• Relief: a transition period, official guidance, and sector-specific regulatory sandboxes that give room for trial and error.
• Industrial policy: continued encouragement of open AI development, the approach that drove wide adoption of Chinese open-weight models abroad.
• Social policy: monitoring AI's effect on employment and displacement, with safety-net protections for affected workers.

What this means for developers and labs

For labs like DeepSeek, Moonshot and Alibaba, the practical questions are about paperwork and timing. We do not know yet whether disclosure will be a legal duty, who receives incident reports, or what counts as a reportable incident.

For teams outside China that use Chinese open-weight models, the thing to watch is whether lifecycle rules change release practices: model cards, safety evaluations or staged releases. We have already tracked how Chinese labs disclose safety information in our post on China AI labs' safety disclosure across 857 releases. Its finding was that disclosure had been thin; a mandatory standard would be a direct change.

The compute side also matters. Wang named computing power as one of the regulated enablers, which ties into China's capacity build-out discussed in China's 24 GW compute versus the US 56 GW.

Policy gavel and scroll illustrating China AI legislation and regulation

How we got here: China's rules so far

This announcement does not start from zero. China has regulated AI through a series of narrower instruments, usually issued by the Cyberspace Administration of China (CAC) together with other ministries.

1. 2022 to 2023: algorithm and deep synthesis rules. Recommendation algorithms and deepfake-style synthetic media came first, with registration and labeling duties.
2. August 2023: generative AI measures. The Interim Measures for the Management of Generative AI Services required providers of public-facing models to meet content and security requirements and to file their models with regulators.
3. 2025: labeling rules and safety frameworks. Rules requiring AI-generated content to be labeled took effect in September 2025, and the national standards body TC260 has published an AI safety governance framework and draft standards, including ones about loss of control.
4. Voluntary industry pledges. Many leading Chinese developers signed voluntary AI safety commitments drafted by an industry alliance.

The difference now is scope and level. Earlier rules targeted outputs, content and filing. A lifecycle standard reaches into training data, compute, evaluation and incident handling, which is closer to how the EU AI Act and the US frontier-model proposals think about the problem. Wang's wording, data, computing power and algorithms, is the same set of inputs that frontier-lab safety frameworks already talk about. These background points come from our general knowledge of public Chinese rule-making and should be checked against the CAC and TC260 sites before you rely on exact dates.

Questions developers are asking

Does this affect me if I only use DeepSeek or Qwen through an API? Possibly in the long run, not today. Duties as described fall on developers and operators in China. API users outside China are not named.

Will open weights keep coming? The stated policy is yes. Bloomberg quotes the government continuing to encourage open development to 2030. That matters because Chinese open-weight models are widely used abroad, as we covered in our look at how cheap Chinese open models change AI infrastructure.

Could disclosure rules reveal more about Chinese models? If incident and vulnerability disclosure becomes mandatory and public, outside researchers could learn far more about failure modes than today. If it is reported only privately to regulators, outsiders may learn little. That detail decides how useful the rule is, and it is not yet public.

Is this about frontier risk or about content control? The reported remarks stress safety, loss of control and cyber risk. Observers will watch whether the standards also carry content-moderation duties, as earlier Chinese rules did. Nothing reported so far settles that.

Why now? Bloomberg points to incidents where AI agents went rogue or boosted cyberattacks. We have seen such cases elsewhere, such as the OpenAI grader that destroyed its own environment. Governments in several countries are responding to the same pattern at about the same time.

How it compares with other governments

Incident reporting is becoming a shared idea. In the United States, a White House order on AI incident disclosure is covered in White House mandates AI incident disclosure, and House Democrats proposed 48-hour and 24-hour reporting windows in the Pallone, Lofgren and Raskin proposal. In Europe, the AI Act commissioner discussed rogue agents in Virkkunen's Reuters interview. India's cyber agency issued its own warning, covered in CERT-In's advisory on AI agents.

Bloomberg also notes that China and the United States agreed in September to try to build a common understanding of AI's benefits and threats, ahead of a summit between the two presidents in Washington. That is a signal, not a treaty.

Claimed vs verified

To check for yourself, watch the NDRC website for the briefing transcript, the Cyberspace Administration of China for draft rules, and the State Council legislative agenda for the bill text.

What to do now

• If you ship or fine-tune Chinese models, record model versions and evaluation results now, so disclosure later is easy.
• If you operate agents that touch Chinese systems, assume incident-disclosure duties may arrive and keep logs.
• If you build agents, runtime guardrails help in every jurisdiction. AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
• Do not treat "standards" as laws yet. Wait for text.

What is still open

• The legal form: a law, regulations, or voluntary national standards.
• The timetable and the length of the transition period.
• Who enforces, and the penalties.
• What counts as a reportable incident or vulnerability.
• How sandboxes will be chosen and which sectors go first.

We will update this post if the NDRC or the Cyberspace Administration publishes text.

Related reading

• China AI labs' safety disclosure, 857 releases
• China 24 GW compute vs US 56 GW
• White House mandates AI incident disclosure
• House Democrats' AI incident report proposal
• EU's Virkkunen on rogue AI agents
• CERT-In advisory on AI agents
• India and rogue AI agents

Details are accurate as of November 3, 2026. The plan is reported intent, not final law.