AI Agent Security Can't Rely on Prompts Alone: Shanghai AI Lab Explores a New Paradigm for Agent Safety Evolution

Shanghai AI Laboratory, together with Fudan University, Shanghai Jiao Tong University, Hong Kong University of Science and Technology, and Zhejiang University, proposed SHE and SafeEvolve to explore a new paradigm for large-model agent safety evolution. SHE starts from complete execution trajectories and decomposes the safety harness into four components—System Prompt, Rule Bank, Safety Memory, and Tool Policy—enabling targeted revisions and validation through trajectory diagnosis. SafeEvolve further writes validated harness safety experience into the policy model, using stages such as Harness-use SFT to strengthen the model's own safety decision-making.