Sina Science and Technology News on the afternoon of October 13, at the 38th National computer Security academic Exchange, Zhou Hongyi, founder of 360 (hereinafter referred to as "360"), revealed the details of 360 model security solutions. He pointed out that the development of the large model has entered the second stage, the focus is to lower the figure, the digital development of the enabling industry, the biggest obstacle is how to create a truly safe and usable large model.
Zhou Hongyi pointed out that the big model is a "double-edged sword" that will bring unprecedented security challenges. he divides security risks into short-term, medium-term and long-term stages. Among them, short-term risk is the most urgent problem to be solved, mainly around the security problems caused by the large model itself, such as network security, data security and generated content security. Medium-term risk mainly refers to man-made malicious applications of large models, such as the increasingly rampant AI face-changing and AI voice-changing fraud. The long-term risk is the security challenge to human survival when more powerful artificial intelligence is born.
Zhou Hongyi says that the safety problems of the basic large model will be "inherited" to the vertical large model. To this end, 360 bowed to enter the Bureau, on the basis of national security capabilities, to create a primary security model "360 brain". Zhou Hongyi said that on the basis of the four security principles of "safe, good, credible and controllable", 360 has built a systematic security solution for large models to ensure the landing of large models in enterprise-level scenarios. He revealed that in order to keep the large model from doing evil, the core of the 360 "good" solution is to develop a safe vertical "small model" to train and monitor the large model to ensure the safety of the output of the big model.
In order to solve the problem of "controllability" of large models, 360360 will launch an Agent framework to give human capabilities to large models, which on the one hand make the large models more powerful, on the other hand, they can also act as constraints on large models. In the face of the blowout of domestic large model security requirements, Zhou Hongyi said that 360's native large model security solutions can be translated and copied to industry partners and customers in need, and empower externally in the form of "360 big model security butler". Provide network and data security, content guardrail, search knowledge enhancement, visual data monitoring and other capabilities. (Wen Meng)