Plimsoll makes AI code sandboxes declare their isolation level
A new Show HN post presents a self-hosted runner that labels every agent-written code execution by isolation tier and refuses jobs when the available tier is too weak.
A visible floor for agent-written code
Plimsoll’s creator has put a self-hosted sandbox for AI-written code in front of the Hacker News community. The pitch is that a developer should be able to ask for a minimum isolation level before an agent’s program runs. Plimsoll labels each result as process, container, kernel or virtual-machine isolation; if the configured provider cannot meet the requested level, the project says it refuses the job before execution. That makes the boundary visible to the calling tool instead of burying it in deployment documentation.
The English README describes request and result hashes that clients can check, plus a separate tool for signing a chain of run records. It also describes an API broker that keeps credentials outside the sandbox and grants agent code access only to specified routes. The creator’s Show HN post says the project can use local and cloud providers and includes Go, Python and TypeScript clients. These are the maker’s descriptions of the design, not an independent security assessment.
The boundary still needs scrutiny
The creator says Plimsoll is pre-1.0, has no outside security audit and had no other users at the time of the showcase. Its quick-start WebAssembly example reports only process-level isolation, which the README explicitly warns is unsuitable for hostile code. The reported tier comes from configuration and startup checks rather than hardware attestation. Teams considering it for untrusted agent output still need to verify the provider and deployment they actually use.