OpenSciFlow workflows can execute local commands, so security is central.

All standards and prototypes are drafts. Do not run untrusted manifests or workflows on sensitive systems.

Please open a private security advisory if available, or contact the maintainers privately before public disclosure.

• Do not execute arbitrary LLM-generated shell commands.
• Use allowlisted plugin commands.
• Restrict paths to approved work directories.
• Require user approval before downloading model weights or containers.
• Record commands and environment metadata for every run.