On 13 July 2026 the National Industrial Information Security Development Research Centre, an institution under the Ministry of Industry and Information Technology, issued a notice launching a co-construction initiative for large-model safety evaluation benchmarks and inviting participants.
According to the notice, the benchmark constructs a panoramic assessment framework around the safety governance needs of real-world large-model applications, covering six core dimensions — content safety, value alignment, robustness, fairness, privacy protection and truthfulness — and forming a systematic end-to-end evaluation system. The evaluation methodology combines automated fuzzing, stress testing and human supervision, and covers five categories comprising 31 specific safety risks.
Co-construction participants include large-model technology companies, industry application organisations, research institutions and safety specialists, who will take part in standard-setting, testing practice and publication of results.
Why this matters
This is not another consultation draft or industry initiative. It signals a clear shift: the regulator is moving from rule-making to capability verification.
The pace of AI regulation over the past two years bears this out:
- August 2023: the Interim Measures for the Administration of Generative AI Services came into force, establishing the basic framework.
- From 2024: algorithm filing and deep synthesis filing became routine.
- 2025: several departments conducted joint special inspections of generative AI.
- April 2026: the Interim Measures for the Administration of Anthropomorphic AI Interaction Services were issued and came into force on 15 July.
- July 2026: Shanghai launched the “Cast Shield, Model Capital” special campaign, and the co-construction of the safety evaluation benchmark followed.
From legislation to enforcement, from filing to evaluation, the regulatory chain is closing. The introduction of the benchmark means that companies will in future need not only a declaration of compliance but proof of it: whether a model is safe will no longer be a matter for the company’s own assertion, but a matter for standardised evaluation.
The six dimensions
The six dimensions are analysed below from the perspective of compliance practice.
1. Content safety
The core concern is whether the model generates unlawful or prohibited information, including violence, pornography, false information and content harmful to national security. Companies need a sound content filtering mechanism, covering the cleaning of training data, moderation of generated content and updating of sensitive-word libraries.
Practical effect: for consumer-facing AI products, content safety is the first line of defence. The benchmark is likely to include adversarial test cases examining the boundaries of content safety under induced prompts.
2. Value alignment
The model’s output is required to accord with mainstream values and public order and good morals, and not to produce discriminatory, extreme or misleading content. This engages the RLHF (reinforcement learning from human feedback) strategy at the training stage and the quality and diversity of the alignment dataset.
Practical effect: value alignment is one of the most technically demanding dimensions. Companies must demonstrate the effectiveness of their alignment strategy rather than rest on a statement that alignment “has been done”.
3. Robustness
The dimension tests the stability of the model under adversarial attack, anomalous input and boundary scenarios, including resistance to jailbreak attacks, tolerance of adversarial samples and consistency of behaviour under extreme input.
Practical effect: several recent AI safety incidents have been connected with inadequate robustness. The benchmark is likely to include automated fuzzing and stress testing that simulate malicious attack scenarios.
4. Fairness
The concern is whether model output exhibits systematic bias across groups, including on grounds of sex, region, age and occupation. Companies are expected to detect bias at the training data stage and to establish fairness monitoring at the output stage.
Practical effect: fairness is particularly sensitive in recruitment, financial services and healthcare. Companies need quantifiable fairness metrics rather than general statements that the model is “non-discriminatory”.
5. Privacy protection
It is prohibited to use users’ chat records and uploaded data for model training without express authorisation. Companies are required to establish data segregation, the capacity to delete user data, and processes for identifying and masking sensitive information.
Practical effect: this is one of the regulatory red lines. The Shanghai “Cast Shield, Model Capital” campaign has expressly required companies to retain data processing logs for not less than six months and to establish data traceability and audit mechanisms. Companies must demonstrate the compliance of their processing rather than rely on a general authorisation in the user agreement.
6. Truthfulness
The model is required to answer accurately within the limits of its capability, not to fabricate facts — the control of hallucination — and to inform the user clearly where it is uncertain. In high-risk fields such as finance and healthcare the requirement is stricter.
Practical effect: hallucination is one of the central technical challenges of current large models. The benchmark may include fact-accuracy test sets for particular fields, such as law, healthcare and finance, testing the professional reliability of the model.
The practical impact on AI companies
Short term (the second half of 2026)
First, compliance costs will rise. Whether or not a company participates in the co-construction, every large-model company will need to self-assess against the six dimensions and prepare for possible evaluation requirements.
Second, the industry will polarise faster. Companies with solid technical foundations and adequate investment in safety will gain a competitive advantage, while those relying on a thin wrapper or with weak safety capability will face a higher compliance threshold and the risk of being cleared from the market.
Third, some application scenarios will be constrained. Until the evaluation standard is settled, the rollout of certain high-risk applications — AI medical consultation, AI financial advice and AI legal services among them — may slow.
Medium and long term (2027 onwards)
First, safety evaluation may become a market-entry requirement. Following the experience of algorithm filing and deep synthesis filing, large-model safety evaluation may in due course be made mandatory, and services that have not passed evaluation may be restricted from launching or operating.
Second, safety compliance will become a competitive feature. Passing the evaluation will serve as an endorsement of trust, comparable to app store review, and a safety label will become an important reference for users in choosing a product.
Third, demand for legal services will grow. The preparation of safety assessment reports, the design of remediation plans and engagement with the regulator all require professional legal support, and AI safety compliance will become a new area of growth for legal practice.
How companies should prepare
Step one: establish the current position
Assess the product as it stands against the six dimensions — content safety, value alignment, robustness, fairness, privacy and truthfulness — and identify the weak points.
Step two: technical remediation
Prepare a technical remediation plan on the basis of that assessment, focusing on the cleaning of training data, content filtering, defence against adversarial attack, segregation and deletion of user data, and control of the hallucination rate.
Step three: institutional arrangements
Establish a safety management system covering the whole of AI development, training, deployment and application, with clear allocation of responsibility and a closed loop of risk identification, monitoring, response and disposal.
Step four: preparation for evaluation
Drawing on the experience of algorithm filing and deep synthesis filing, prepare safety assessment materials in advance, including but not limited to the technical architecture description, safety strategy documents, test reports, the user agreement and data processing records.
Step five: professional support
A professional legal team should be engaged to assist with the compliance assessment, the remediation plan and engagement with the regulator. Safety evaluation sits at the intersection of technical and legal questions and requires a team with experience in the AI sector.
On participation in the co-construction
The deadline for applications to participate was 15 July 2026. Applicants include large-model technology companies, industry application organisations, research institutions and safety specialists.
Participants are entitled to take part in standard-setting, testing practice and publication of results, and to contribute to a benchmark with authority and industrial influence.
For AI companies with technical depth and industry standing, participation is both a responsibility and an opportunity: involvement at the standard-setting stage assists early understanding of regulatory expectations and reduces subsequent compliance costs.
Smaller companies that do not presently meet the criteria are advised to follow the formal publication of the benchmark and to self-assess and remediate against it in advance.
This article is prepared on the basis of the notice of the National Industrial Information Security Development Research Centre of 13 July 2026. For assistance with the compliance implications of the large-model safety evaluation benchmark, please contact Geason Law Offices.