360 Participates as a Founding Member of the "Self-Discipline Convention on Agent Development" to Build a Trusted Agent Ecosystem

On November 4, at the "Global Cooperation and Innovation on Mutual Trust, Interconnection, and Interoperability of Agents" forum of the 2026 World Internet Conference Wuzhen Summit, the China Academy of Information and Communications Technology, 360, Huawei, and other enterprises and institutions jointly released the "Self-Discipline Convention on Agent Development" (hereinafter referred to as the "Self-Discipline Convention"). As one of the core initiating units, 360 deeply participated in the drafting and release of the Convention.

"Global Cooperation and Innovation on Mutual Trust, Interconnection, and Interoperability of Agents" Forum

Currently, agents are accelerating from "single-point applications" toward "multi-agent collaboration." However, when agents interact across platforms and systems, security risks such as their own permission overreach, data leakage, and loss of behavioral control have become core pain points hindering their large-scale deployment. To implement the national task deployment of "strengthening industry self-discipline," the Self-Discipline Convention released this time proposes eight self-discipline measures, focusing on key issues such as improving the agent trust system, enhancing security protection capabilities, and improving governance mechanisms.

Article 4 of the Self-Discipline Convention sets clear requirements for agent security protection, which aligns with 360's technical approach in recent years of "using models to govern models." The Convention proposes to "improve technical systems for forward-looking risk assessment, operational monitoring, behavior control, and anomaly circuit-breaking." In response, 360's solution is to use AI to govern AI. At present, 360 has built two lines of defense, "large model security" and "agent security," specifically addressing the endogenous risks of AI systems and the unauthorized operations and loss of behavioral control caused by external attacks.

Particularly in terms of agent operational security, 360 has launched a dedicated agent security protection and management system. The system covers asset discovery, behavior control, threat detection, incident response, and loss recovery. During operation, the system can not only identify and manage agents and skill plugins, but also implement strict "runtime control" over file access, command execution, and more. Combined with security large models, the system can promptly detect risks such as prompt injection and malicious skills, and supports complete tracing of anomalous events and rapid recovery of the environment, enabling agent environments to truly move from "visible" to "manageable, traceable, and recoverable."

In specific product practices, 360 has deeply embedded security capabilities into the application layer. For individual users, 360's Nano Work runs agents in an isolated cloud environment, without access to users' local data unless authorized, and can instantly rebuild the isolated environment to control risks in the event of anomalies; for government and enterprise customers, 360's sovereign AI operating system has been deployed in some enterprises, enabling them to build exclusive agents while firmly retaining ultimate control over data, knowledge, and security governance, achieving the true implementation of "secure agents."

Industry insiders point out that the foundation of the healthy development of agents lies in security and mutual trust. The deep participation of enterprises such as 360 will help drive the Self-Discipline Convention from industry consensus to technical implementation, providing a solid foundation for building an open, trustworthy, and secure agent ecosystem.

Leiphone Leiphone