China works on AI safety benchmark as regulators target large model risks

State-led initiative addresses AI concerns such as data leaks and hallucinations, as Beijing joins global efforts to enhance oversight of generative models

China’s Ministry of Industry and Information Technology (MIIT) has started building a safety benchmark to evaluate artificial intelligence models, as regulators in the United States and Europe strengthen oversight of AI security.

The MIIT-led National Industrial Information Security Development Research Centre is now recruiting companies and experts to co-build the benchmark, with applications due on Tuesday, according to a notice published on Monday.

The institute said that current frameworks fail to meet complex safety-governance needs, requiring a standardised testing platform to support industrial compliance.

The new benchmark will evaluate generative AI across six core dimensions: content safety, value alignment, robustness, fairness, privacy protection and trustworthiness, according to the notice. A hybrid benchmarking methodology will be designed to explicitly cover 31 specific safety risks across five major categories.

The system will combine automated fuzzing and stress testing with human oversight to control hallucination rates and data leaks, as well as curb “jailbreak” attacks – malicious prompt engineering techniques designed to bypass the safety barriers of large language models (LLMs).

This state-led technical approach comes as global regulators have been putting extra emphasis on AI safety as the technology advances.

In June, the White House issued an executive order calling for “a classified benchmarking process to assess the advanced cyber capabilities of AI models”.

It followed the April release of Anthropic’s Mythos – the company’s most powerful model to date – which has been withheld from general use due to its unprecedented capacity to spot and exploit cybersecurity vulnerabilities.

Meanwhile, the European Union is executing a strictly codified legal framework under the EU AI Act, which mandates training-data transparency for general-purpose AI models, backed by statutory financial penalties.

While violations face fines of up to 3 per cent of a company’s global turnover, non-compliance involving prohibited AI practices could trigger peak penalties of up to 7 per cent.

Source: scmp.com