Model Laboratory treats experiment files, model source, local attachments, interpreter output, and extension declarations as potentially untrusted inputs.

Security maintenance currently targets the latest 1.17.x source and release line. Historical formats remain covered by the compatibility and safe-open rules documented in the repository.

Use the repository's private security-advisory facility when it becomes available. Include:

• the affected version and platform;
• the relevant file, parser, IPC action, bundle member, capability, or workflow;
• a minimal reproduction or proof of concept;
• expected and observed behavior; and
• any known effect on confidentiality, integrity, availability, or reproducibility.

Please use a public issue only after a coordinated fix or disclosure has been prepared.

• Opening a .mlab performs bounded integrity and schema inspection without scientific execution.
• Experiment bundles declare capability requirements but do not carry executable extension code.
• Reproduction and capability execution are explicit actions using locally installed code.
• Interpreter output is parsed, bounded, schema-validated, compiled, and presented for acceptance.
• Local reference attachments use bounded extraction and content-addressed provenance.
• Workload estimation and capability settings validation occur before a runner starts.
• Committed experiment receipts bind the complete state and require their checksum for external execution interfaces.
• Network access is limited to explicit optional runtime operations such as obtaining a configured local model through the user's independently installed provider.

Release builds should be produced from a tagged source state, preserve dependency lockfiles, generate build-identity evidence, and publish cryptographic checksums. Third-party dependencies and vendored assets should be reviewed against the resolved release locks.