golden-dataset
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEDYNAMIC_CONTEXT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_CONTEXT_INJECTION]: The skill utilizes platform-specific !command syntax in scripts/backup-golden-dataset.md to execute shell commands during the loading phase. These include date, pwd, python --version, and find commands scoped to identify dataset directories. While used for environment auto-detection, this involves silent execution of local shell commands.
- [INDIRECT_PROMPT_INJECTION]: The curation workflow (rules/curation-add-workflow.md) processes untrusted data from external URLs via WebFetch. This ingestion point exposes the agent to adversarial content. Mandatory Evidence Chain: (1) Ingestion: WebFetch from URLs; (2) Boundaries: None specified in prompt templates; (3) Capabilities: Read, Glob, WebFetch, WebSearch, and local command execution; (4) Sanitization: None defined for fetched content.
- [COMMAND_EXECUTION]: Instructions and rule files (rules/management-versioning.md, rules/validation-regression.md) guide the agent to perform local command-line operations using python and poetry for dataset backups, integrity verification, and regression testing.
Audit Metadata