Anthropic Introduces sandbox-runtime: OS-Level Sandboxing for AI
Anthropic launches sandbox-runtime, an open-source OS-level sandboxing tool powering Claude Code using Seatbelt on macOS and bubblewrap on Linux without containers.
Stock photo for illustration only, not from the actual event
- Anthropic launches open-source sandbox-runtime (srt)
- It serves as the OS-level sandbox behind Claude Code
- Restricts file and network access for AI agents, MCP servers, and CLI tools
- Utilizes Seatbelt on macOS and bubblewrap on Linux without containers
Anthropic has introduced a new open-source tool named sandbox-runtime, or srt, designed to provide OS-level sandboxing. This runtime operates as the underlying security mechanism behind Claude Code to enhance safety during operations.
The primary function of srt is to restrict file and network access for various high-risk components, including AI agents, MCP servers, and CLI tools, preventing them from accessing unauthorized system resources.
Stock photo for illustration only, not from the actual event
OS-level sandboxing differs from traditional containerization technologies like Docker by leveraging native security features built directly into the operating system. This approach offers faster execution and lower resource overhead, making it ideal for AI agents that automatically execute commands or write code, as limiting their file and network scope mitigates the risk of executing harmful operations.
Under the hood, srt relies on established operating system security frameworks, implementing Seatbelt when running on macOS and bubblewrap when operating on Linux platforms.
A key advantage of this architecture is that it operates entirely without containers, reducing deployment complexity and boosting overall execution efficiency.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment