CNCERT reports 2026 large model security crowdsourcing test results: 873 vulnerabilities found

The 2026 AI large model security crowdsourcing test results were released on September 15 at the National Cybersecurity Awareness Week. 2,467 white-hat hackers tested 54 large models and agent application products from 25 domestic AI vendors, discovering a total of 873 security vulnerabilities, including 608 vulnerabilities unique to large models and agent applications, and 265 traditional vulnerabilities (30.4%). The report lists six typical risk categories: prompt injection vulnerabilities are widespread, allowing models to be induced to execute malicious instructions pre-embedded in web pages or documents; information leakage vulnerabilities are frequent, with one large model having hardcoded and leaked API keys, exposing API keys for 239 models on a cloud platform; improper output handling vulnerabilities can lead to remote command execution and privilege escalation; agents have issues such as proxy identity and permission abuse, proxy goal hijacking, and unintended code execution, for example, an agent's sandbox process management service API interface was exposed, allowing arbitrary system commands to be executed with administrator privileges. Handling recommendations include treating model outputs as untrusted input, least privilege and sandbox isolation, input/output filtering, sensitive data masking, vendor audits, and promoting a closed-loop vulnerability handling process combining crowdsourced testing and intelligent monitoring.