On May 9, 2026, the Beijing Academy of Artificial Intelligence, working in tandem with esteemed institutions like Peking University and Beihang University, officially launched the FlagSafe Large Model Security Platform. This platform zeroes in on three core areas: red teaming exercises (proactive security testing), blue teaming defense (reactive security measures), and white box analysis (transparent security evaluation). By doing so, it constructs a comprehensive closed-loop system that encompasses risk identification, defense strategies, and mechanism clarification.

In light of the multifaceted and cross-scenario composite risks that large models encounter, the platform delineates five security red lines—non-negotiable standards for ensuring model safety. Additionally, it has initially crafted a security capability matrix that spans three major dimensions, offering a structured approach to security assessment. Its overarching goal is to furnish the industry with a robust security resource pool and evaluation framework that seamlessly integrates "assessment, defense, and explainability," thereby fostering a more secure and trustworthy AI ecosystem.