Beijing, March 11, 2025 – In a surprising turn of events, AI startup Manus has revealed that its foundation models are Claude Sonnet and a fine-tuned version of Alibaba’s Qwen, following the discovery of its sandbox code by internet users. The revelation comes after a user, identified as @jianxliao, successfully accessed the Manus sandbox code by exploring files within the /opt/.manus/ directory.
The discovery, initially reported by Caijing (First Finance), highlights the increasing transparency – or lack thereof – in the rapidly evolving AI landscape.
The user’s findings revealed that Manus leverages Claude Sonnet, along with 29 tools built upon it, but does not utilize Claude’s multi-agent capabilities. Furthermore, the company employs the open-source code from Browser Use as the foundation for its Computer Use functionality. Computer Use, a feature introduced by Anthropic as part of its Claude 3.5 Sonnet model, enables the AI to interact with computers in a human-like manner, including tasks such as viewing screens, moving cursors, clicking buttons, and entering text.
The /opt/ directory is a standard system directory commonly used to store third-party software or optional packages. A sandbox, in this context, is an isolated environment used for testing, executing, or running code while restricting its access to system resources. This isolation is crucial for security and stability.
The user’s simple request yielded the Manus runtime code, prompting the public unveiling of its underlying technology.
In response to this passive disclosure, Manus co-founder Ji Yichao stated that the leak of the sandbox code was not an accident or vulnerability, but rather an intentional design feature. According to Ji, each conversation operates within an independent sandbox environment, completely isolated from other sessions. Users can directly access the sandbox through the Manus interface. The code within the sandbox is solely responsible for receiving commands from the Agent, and the Multi-Agent implementation is a core characteristic of Manus.
The sandbox’s primary function is to provide a secure and controlled environment, preventing code from causing potential harm to the host system or other user data, Ji elaborated. Some sandbox environments may offer developer modes or debugging features, allowing users to inspect the code and understand the Agent’s decision-making process.
This incident underscores the ongoing debate surrounding transparency and explainability in AI development. While Manus claims the code accessibility is by design, the initial discovery by an external user raises questions about the company’s initial disclosure strategy.
The incident also highlights the importance of understanding the role of sandboxes in AI development and deployment. Sandboxes are crucial for ensuring the safety and security of AI systems, particularly as they become more integrated into everyday life.
The use of Claude Sonnet and Alibaba’s Qwen as foundation models positions Manus within a competitive landscape. While Claude Sonnet is a well-regarded model known for its performance and capabilities, the integration of a fine-tuned Qwen model suggests a strategic effort to leverage regional expertise and potentially optimize for specific tasks or datasets.
The future will reveal how Manus leverages these foundation models and its unique sandbox environment to differentiate itself in the rapidly evolving AI market.
References:
- Lü, Q. (2025, March 10). 被网友获取运行代码后，Manus终于披露基座模型是哪家 [Manus finally discloses the base model after the running code was obtained by netizens]. Caijing. Retrieved from [Insert Original URL Here if Available]
Views: 0