2024 Large Model Security Practice Whitepaper Unveiled at the World AI Conference

On July 5, a whitepaper titled "Large Model Security Practice (2024)"—co‑authored by Tsinghua University, Zhongguancun Laboratory, Ant Group and others—was officially released at the 2024 World Artificial Intelligence Conference.

The whitepaper systematically introduces an overall security practice framework, offering technical implementation solutions across safety, reliability, and controllability dimensions, along with case studies in finance, healthcare, and government, and a "five‑dimensional integrated" governance model to support high‑quality industry development.

Large‑model technology is becoming a key driver of societal progress, yet its growing capabilities raise unprecedented challenges in security, reliability, and controllability, including data leakage, value alignment, hallucinations, and risks in data, model, algorithm, and hardware environments.

The proposed framework emphasizes three core pillars—security (protecting data, model, system, content, cognition, and ethics), reliability (ensuring consistent, accurate, and truthful outputs), and controllability (enabling human understanding and intervention)—to enhance robustness, explainability, fairness, privacy, and value alignment.

The whitepaper also highlights the importance of security testing and defense technologies, noting current evaluations focus on content scenarios while future challenges involve complex agent architectures and AGI, calling for standards and collaborative efforts among governments, academia, and industry.

As a concrete example, the whitepaper showcases Ant Group’s in‑house solution "Ant Tianjian," a comprehensive large‑model security assessment and defense product already adopted by over 20 external institutions in finance, healthcare, and government.

In the financial sector, Ant AI Assistant "ZhiXiaoBao" secures large‑model training, inference risk control, comprehensive risk assessment, and user interaction risk management, incorporating consistency checks and financial value alignment.

In healthcare, Shanghai First People’s Hospital employs Ant Tianjian’s safety‑prevention technology to eliminate critical risks, ensuring generated content meets medical safety and professionalism standards while protecting privacy.

In the government domain, the "GanFutong" AI assistant implements on‑device safety measures, large‑scale intent recognition, intelligent questioning, and a robust security shield covering hundreds of content‑generation risks, capable of handling up to 500,000 concurrent attacks.

Tsinghua University associate professor Li Qi notes that large‑model security is an emerging field still in its infancy, with many enterprises adapting traditional data, information, and system security expertise to this new context.

Ant Group’s chief scientist Wang Weiqiang, in his keynote "Exploring Trustworthy Practices for Large‑Model Applications," emphasized that enhancing security, reliability, and controllability within the existing trustworthy AI governance framework is essential for sustainable AI development.

The whitepaper concludes with a five‑dimensional, multi‑stakeholder governance framework—encompassing government regulation, ecosystem cultivation, corporate self‑discipline, talent development, and testing verification—to promote a secure and sustainable large‑model ecosystem.