China's MIIT builds AI safety benchmark targeting 31 LLM risks
Synopsis
China's MIIT is co-building a national AI safety benchmark that will test large language models across 31 specific risks — including jailbreaks, hallucinations, and data leaks — marking Beijing's shift from policy rhetoric to measurable technical enforcement.
Key Takeaways
China's MIIT is developing a national AI safety benchmark through the National Industrial Information Security Development Research Centre , with expert recruitment closing Tuesday, 15 July 2026 .
The benchmark will evaluate generative AI across six core dimensions : content safety, value alignment, robustness, fairness, privacy protection, and trustworthiness.
A hybrid methodology will cover 31 specific safety risks across five major categories , combining automated fuzzing, stress testing, and human oversight.
The system is specifically designed to counter 'jailbreak' attacks — prompt engineering techniques that bypass LLM safety barriers — as well as hallucination and data-leak risks.
The initiative runs parallel to tightening AI oversight in the United States and Europe , signalling a global race to define enforceable AI safety standards.
China's Ministry of Industry and Information Technology (MIIT) has launched an initiative to construct a national AI safety benchmark for evaluating large language models, as regulators worldwide intensify scrutiny of generative AI risks. The MIIT-led National Industrial Information Security Development Research Centre is actively recruiting companies and experts to co-develop the framework, with applications closing on Tuesday, 15 July 2026.
What the benchmark covers
According to the official notice published on Monday, 14 July 2026, the new benchmark will evaluate generative AI across six core dimensions: content safety, value alignment, robustness, fairness, privacy protection, and trustworthiness. A hybrid benchmarking methodology will explicitly address 31 specific safety risks spanning five major categories, the notice stated.
The institute acknowledged that existing frameworks 'fail to meet complex safety-governance needs,' citing the urgency for a standardised testing platform to support industrial compliance across the sector.
How the system works
The proposed framework will combine automated fuzzing and stress testing with human oversight to control hallucination rates and prevent data leaks. It will also specifically target 'jailbreak' attacks — malicious prompt engineering techniques designed to bypass the safety barriers of large language models (LLMs).
This multi-layered approach signals that Beijing is moving beyond high-level policy pronouncements toward technical, measurable compliance infrastructure for AI developers.
The competitive backdrop
The MIIT initiative arrives as regulators in the United States and Europe have been strengthening their own AI oversight mechanisms. The EU AI Act and various US executive actions have placed LLM safety at the centre of the global policy agenda, creating pressure on China to establish comparable — and sovereign — evaluation standards.
Chinese AI developers, who operate under existing interim generative AI regulations introduced in 2023, would likely be required to align with the new benchmark once formalised, affecting a broad swath of the domestic industry.
Why it matters
A state-backed, standardised safety benchmark gives Chinese regulators a concrete technical instrument to enforce compliance rather than relying solely on self-reporting by AI companies. It also positions China to export its governance model to markets in Asia, Africa, and the Middle East that are still shaping their own AI regulatory frameworks.
Industry analysts have noted that the move reflects a broader global pattern: governments racing to define what 'safe AI' means in measurable terms before the technology outpaces their oversight capacity.
What's next
With the recruitment window closing imminently, the pace at which the National Industrial Information Security Development Research Centre assembles its expert coalition will signal how seriously Beijing intends to fast-track the benchmark's deployment. Developers of frontier models operating in China should watch for draft testing protocols that could reshape compliance timelines across the sector.
Point of View
Once formalised, becomes a template that Belt and Road partner nations may adopt by default, quietly extending China's normative influence over global AI governance. The focus on 'jailbreak' attacks and hallucination control also reveals that Beijing is acutely aware of the reputational and security risks posed by its own frontier models in high-stakes deployments. Developers with dual-market ambitions — serving both Chinese and international customers — will face the growing compliance cost of satisfying two increasingly divergent safety regimes.
NationPress
13 Jul 2026
Frequently Asked Questions
What is China's new AI safety benchmark?
China's new AI safety benchmark is a state-led evaluation framework being developed by the MIIT -led National Industrial Information Security Development Research Centre . It will test generative AI models across six core dimensions and 31 specific safety risks , covering threats such as jailbreak attacks, hallucinations, and data leaks.
Which government body is leading China's AI safety benchmark?
The initiative is led by China's Ministry of Industry and Information Technology (MIIT) through its National Industrial Information Security Development Research Centre . The centre published a notice on Monday, 14 July 2026 calling for companies and experts to join the co-development effort.
What risks will China's AI benchmark evaluate?
The benchmark will cover 31 specific safety risks across five major categories , including content safety, value alignment, robustness, fairness, privacy protection, and trustworthiness. It will specifically address 'jailbreak' prompt-engineering attacks, hallucination rates, and data leakage from large language models.
Why is China building its own AI safety standard?
According to the official notice, existing frameworks 'fail to meet complex safety-governance needs,' necessitating a standardised testing platform for industrial compliance. The move also positions China to establish sovereign AI governance norms as the US and EU advance their own oversight regimes.
Who will be affected by China's AI safety benchmark?
Chinese AI developers — particularly those building or deploying generative AI and large language models domestically — are the primary stakeholders. Once formalised, the benchmark could become a mandatory compliance requirement, affecting companies across China's fast-growing AI sector.
