Co-Construction Action for Large Model Safety Evaluation Benchmarks Officially Launched; Mininglamp Technology (2718.HK) Has Already Planned Ahead with Web Agent Comprehensive Evaluation Benchmark

On July 13, 2026, news from the official WeChat account of the China Industrial Control Systems Cyber Emergency Response Team (CICS-CERT) showed that the Co-Construction Action for Large Model Safety Evaluation Benchmarks has been officially launched.

It is understood that CICS-CERT, relying on its technical accumulation and practical experience in the field of large model safety evaluation, is constructing a large model safety evaluation benchmark. It has now begun recruiting co-construction partners for the benchmark to build a systematic, multi-dimensional artificial intelligence safety evaluation system.

According to the official WeChat article, the large model safety evaluation benchmark under construction builds a panoramic evaluation framework around the safety governance needs in real-world applications of large models. It covers six core safety dimensions: content safety, value alignment, robustness and stability, fairness and impartiality, privacy protection, and authenticity and trustworthiness, thereby forming a systematic, full-link evaluation system.

In addition, CICS-CERT will invite large model technology enterprises, industry application organizations, scientific research institutes, and experts in the safety domain to deeply participate in the co-construction of the system. Deeply involved entities will have the opportunity to become co-construction partners for the large model safety evaluation benchmark, participate deeply in standard formulation, testing practices, and achievement releases, and jointly create an AI safety evaluation benchmark with industry authority and industrial influence.

Within the large model evaluation system, it is worth mentioning that the achievements related to WebRetriever, a self-developed comprehensive evaluation benchmark for large-scale Web Agents by Mininglamp Technology (2718.HK), a Hong Kong-listed large model concept stock, were recently officially accepted by the European Conference on Computer Vision (ECCV 2026), a top-tier international computer vision conference. The benchmark covers 800 real online websites and 1,550 tasks, spanning multiple industries including e-commerce, finance, healthcare, government affairs, and education.

Public data shows that WebRetriever also quantifies for the first time the real capability gap of agents, establishing an application evaluation standard for the agent domain. This evaluation benchmark places Agents in an environment closer to the real internet. Through large-scale task testing, it observes their comprehensive performance in webpage understanding, path planning, interactive execution, and ultimate task completion.

WebRetriever addresses the key limitations of previous works across three aspects: dataset scale and diversity, the reliability of automatic evaluation, and deployment-oriented evaluation protocols.

Industry insiders pointed out that detecting capability shortfalls in Agents through testing can subsequently drive improvements in models, tool invocations, and task planning capabilities. This makes the evaluation system not merely a tool for measurement, but potentially a vital infrastructure for the continuous evolution of Agents. From an industrial perspective, the significance of this type of evaluation system is not just to "score" Agents, but more importantly, to help the industry identify directions for technical optimization, allowing enterprises to judge how far different Agent products are from actual applications.

Zheshang Securities pointed out that the continuous iteration of AI large model capabilities is driving the logic of the network security industry to shift from "AI disrupting security" to "AI amplifying security demands." The security increments brought by AI are not a single software module expansion, but a systematic upgrade running through computing power infrastructure, Agent identity governance, and security operations platforms. The risks that enterprises face regarding identities, data, traffic, and model invocations are rising simultaneously, which in turn is expected to strengthen the willingness to pay for security and fortify the platform moats of leading vendors. If domestic vendors can form platform-based delivery capabilities in directions such as cloud security, Zero Trust, identity security, AI security operations, data security, and security managed services, they are expected to benefit from the rising value of security and the accelerated adoption of localization in the AI era.

NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.