autoresearchclaw-sandboxed-experiment-execution-with-ast-validation

AutoResearchClaw runs all generated experiment code in Docker-isolated sandboxes with AST validation before execution and NaN/Inf fast-fail guards. Code is validated at the AST level to detect unsafe patterns before runtime. The Oracle’s AutoResearch skill has no sandboxing — all experiment code runs in the host environment. This is a security and reproducibility gap for any code-generating research workflows.