AI vs AI: Agent Runtime Security That Gets Stronger with Every Use

Duan Haotong, a Tencent security expert with 12 years of experience, will present at the AICon 2026 conference in Shenzhen on August 21-22. His talk, titled 'Using AI to Fight AI: Building an Agent Runtime Security Line that Gets Stronger with Use,' addresses the emerging risks of prompt injection, intent deviation, tool abuse, and data leakage as agents enter enterprise core scenarios. He argues that traditional rule-based security fails against semantic attacks and non-deterministic behaviors. The proposed solution adopts a first-principles approach with AI-native security, covering pre-incident prompt injection detection (from keyword matching to semantic understanding), during-incident intent deviation detection and agent sandboxing for isolated tool execution, and post-incident data leak prevention via outbound protection and small model fine-tuning. The system relies on a self-iterating feedback loop using real production data. The conference also features ten forums on AI infrastructure, inference engineering, robotic AGI, and agent security, with over 50 experts from Tencent, Alibaba, Huawei, Baidu, and Ant Group.