In-Process Linux Sandboxing backend for agent-sandbox via mazewall

[Pilleo]

Hi everyone,

Currently developers have to choose between the performance and simplicity of LocalSandbox and the latency overhead of Docker sandboxing.

I am developing an open-source library called Mazewall (https://github.com/Pilleo/mazewall) that introduces a middle ground: In-process, kernel-enforced thread-level sandboxing.

Using the modern JDK Foreign Function & Memory (FFM) API, it applies native Linux Seccomp(Like Docker) and Landlock primitives straight to specific JVM thread pools (like virtual thread executors) from inside the running application—without requiring root privileges or native C dependencies.

Why this fits the agent-sandbox abstraction:

The Performance of Core: It runs inside the JVM, meaning zero container cold-start delay, zero heavy infrastructure management, and sub-millisecond execution overhead.
The Isolation of Docker: If an agent tool-calling loop gets compromised or attempts path traversal, the Linux kernel natively returns EPERM at the thread level, throwing a containment exception.

The project is currently in a functional PoC/Pre-Alpha state. I’d love to get thoughts from the maintainers on whether this architecture makes sense as a native, high-performance Linux backend provider (agent-sandbox-landlock) for the Spring AI ecosystem.

Thanks!