Since its launch, the general AI intelligent product Manus has quickly become the focus of the technology industry. Its powerful functions and innovative design have attracted the attention of a large number of users, and there has even been a phenomenon of rushing to buy invitation codes. As the popularity of the product continues to rise, users have also developed a strong interest in the technical architecture and implementation methods behind Manus.
While Manus sparked widespread discussion, some technical teams began to try to replicate its features in an attempt to uncover its mystery. However, what really caused a sensation was a cracking of the Manus system by a user named jian. Through simple commands, Jian successfully obtained the file in the "/opt/.manus/" directory, revealing some of the core information and running code of Manus. This incident not only exposed Manus' technical details to the public eye, but also triggered widespread discussion on the safety of AI products.
According to what jian posted, Manus is not a completely independent model, but is built on Claude Sonnet. In addition, Manus is equipped with 29 tools to assist in task execution, although there is currently no multi-agent function. It is worth noting that Manus also uses an open source project called browser_use, although the relevant code may have undergone some degree of obfuscation. These findings give people a deeper understanding of Manus' technical architecture.
In the face of this incident, the Manus team responded quickly. Co-founder and chief scientist Ji Yichao said that Manus's sandbox system is open, and each session has an independent sandbox environment, completely isolated from each other. He further explained that the code in the sandbox is only used to receive commands, so it is only mildly obfuscated. Ji Yichao also emphasized that Manus' tool design is not mysterious, and the overall architecture is similar to common academic methods. In addition, Manus does use some open source code and promises to open source more content in the future.
When talking about the basic model used by Manus, Ji Yichao revealed that the team mainly used Claude and different Qwen fine-tuning versions. In the early stages of development, they obtained the Claude 3.5Sonnet v1 version and are currently testing Claude 3.7 internally and are looking forward to its update potential. This information further reveals Manus's strategy and future development direction in technology selection.
This incident not only made Manus' technical details more transparent, but also triggered in-depth discussions on AI product security and open source strategies. With the rapid development of AI technology, how to find a balance between innovation and security will become an important topic in future AI product development.