Show HN: Plimsoll, a self-hosted sandbox for code your AI agent writes (github.com)

Plimsoll, a new self-hosted sandbox solution, allows AI agents to run untrusted code securely by classifying the isolation strength of each execution into four tiers: process, container, kernel, and virtual machine (VM). The significance of this innovation lies in its ability to provide rigorous boundaries for executing AI-generated code while ensuring that sensitive credentials are protected. Users can set a minimum isolation strength for code execution, guaranteeing that if the sandbox does not meet specified security requirements, the execution is aborted without risk. Technically, Plimsoll incorporates mechanisms to validate and attest to the safety of code execution, generating SHA-256 hashes for each run to verify consistency and outcomes. The sandbox framework also allows users to design custom "exams" for AI-generated algorithms using predefined scenarios, without exposing sensitive APIs or network access. This design can enhance safety and trust in deploying AI systems by verifying that the code behaves as expected before integrating it into broader applications. With Plimsoll, developers can confidently explore the capabilities of AI without compromising security or stability.