augustus-train
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill provides Python examples (
fit.py,infer.py) that utilizejoblibandpicklefor model serialization. The instructions include a clear security warning stating thatjoblib/pickleloading can execute code and advise users to only load trusted, verified artifacts and validate bundle digests before deserialization. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external records for model training and inference, creating a surface for indirect prompt injection.
- Ingestion points: Data is ingested via JSONL files in
fit.pyandinfer.py, and the general workflow is described inreferences/data-and-splits.md. - Boundary markers: The skill recommends the use of explicit annotation guides and adjudication processes to verify label integrity.
- Capability inventory: The skill performs model training using
scikit-learnand inference viajoblib. It also utilizes companion shell scripts for diagnostics. - Sanitization: Instructions in
references/data-and-splits.mdadvise redacting private fields and removing identifiers from data before processing. - [EXTERNAL_DOWNLOADS]: The skill refers to installing the
scikit-learnpackage via pip and provides links to research notes hosted in the author's official GitHub repository. These references are consistent with the identified vendor and do not involve untrusted or risky sources. - [COMMAND_EXECUTION]: The skill provides documentation for executing several local diagnostic scripts (
policy_receipt.py,provenance_gate.py,overlap_audit.py, etc.). These scripts are intended for local development workflows and operate on user-provided data manifests.
Audit Metadata