CNCERT Publishes 2026 AI Large Model Security Crowdsourced Test Results
CNCERT released 2026 crowdsourced test results on 15 September after 2,467 white-hat testers examined 54 domestic large model and agent products. The campaign recorded 873 vulnerabilities, 608 of them (70%) specific to large models and agent applications.
Results released on 15 September
On 15 September, results of the 2026 artificial intelligence large model security crowdsourced test were released at the Artificial Intelligence Security Governance Forum during National Cybersecurity Publicity Week. The campaign mobilized 2,467 white-hat testers against 54 domestic large model and agent application products. Organizers published outstanding products, outstanding white hats, outstanding test teams, and typical vulnerability risks across three tracks: open-source large models, large model applications, and CLAW-class agent applications.
Outstanding products
The following products had fewer vulnerability findings and no critical or high-severity vulnerabilities.
Open-source large models (in no particular order)
| Outstanding product (open-source large model) | | --- | | Baidu ERNIE-4.5 | | MiniMax MiniMax-M2.1 | | Tencent Hy3 | | DeepSeek DeepSeek-V4 | | Zhipu GLM-4.7-Flash | | ModelBest MiniCPM-o 4.5 | | SenseTime SenseNova-MARS-32B | | Alibaba Cloud Qwen3.5-35B-A3B | | StepFun Step-3.5-Flash |
Large model applications (in no particular order)
| Outstanding product (large model application) | | --- | | Huawei Xiaoyi AI Assistant | | Kuaishou Kling AI | | Xiaomi Xiaoai AI Assistant | | Volcengine Jimeng AI | | Tencent Yuanbao | | Ant Group Ant Afu | | GTCOM (Wuhan) Renduo large model | | DeepSeek Chat | | Volcengine Doubao |
CLAW-class agent applications (in no particular order)
| Outstanding product (CLAW-class agent application) | | --- | | Volcengine Arkclaw | | Moonshot Kimi Claw | | Tencent WorkBuddy | | China Mobile MobileClaw | | Alibaba Cloud QwenPaw | | Huawei Xiaoyi Claw | | StepFun StepClaw | | Alibaba Cloud JVS Claw | | Tencent Qclaw |
Outstanding white hats and teams
Test targets were mainly live AI large model and agent products running under vendor defenses. Organizers credited testers who found high-severity or critical issues under those conditions.
Outstanding white hats (name or handle)
Kyoka, Gu Jiawei, Wen Yuhao, Li Qiang, Liu Junfeng, Zhang Haodong, Yang Pengju, Kn19ht, Li Yanshuo, Wang Shijie.
Outstanding test teams
| Team | Organization | | --- | --- | | 0x300 | Beijing Chaitin Technology Co., Ltd. | | Zhizhi | Beijing NSFOCUS Technology Co., Ltd. | | SecProTech | Hebei SecProTech Computer Consulting Service Co., Ltd. | | HUBUMARS | Hubei University | | Watcher | Beijing Jinyue Wisdom Technology Co., Ltd. | | Wenhai | University of Science and Technology of China | | CNCERT Beijing Branch | CNCERT Beijing Branch | | ChaMD5 | not listed | | Doxs | Shanghai Doxs Information Technology Co., Ltd. | | China Telecom eSurfing Cloud | China Telecom eSurfing Cloud |
Typical vulnerability risks
The campaign found 873 security vulnerabilities in total. Of those, 608 (70%) were specific to large models and agent applications. Organizers said the share shows that current products face a large volume of AI-specific risks in addition to traditional software flaws.
Typical risks reported for open-source large models and large model applications: prompt injection remains the most common class; information disclosure is frequent; some products have improper output handling with serious impact.
Typical risks reported for CLAW-class agent applications: agent identity and privilege abuse is frequent; agent goal hijacking is hard to spot and can produce unintended high-risk operations; some products have unintended code execution with serious impact.
Organizers stated that these products generally have large user bases and high usage, and that abuse of the listed classes would have a relatively serious effect on the domestic large model and agent application ecosystem.
Organizers and follow-up
The campaign was guided by the Cyberspace Administration of China Cybersecurity Coordination Bureau and the National Development and Reform Commission High Technology Department. It was hosted by the National Computer Network Emergency Response Technical Team/Coordination Center of China (CNCERT). The cybersecurity crowdsourcing platform and the National Cybersecurity Talent and Innovation Base provided technical and environment support. Shanghai Intelligent Computing Technology Co., Ltd. and Jingneng Digital Industry Co., Ltd. provided compute support.
Organizers listed four follow-up items: keep improving protection for AI product vulnerability classes; use AI to support crowdsourced testing in a human-machine collaborative mode; use the crowdsourcing mechanism to gather white-hat experts; strengthen monitoring and early warning, and use CNVD to build an AI vulnerability sharing mechanism.