legal-agentic-build
Pass
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill outlines architectures for agents that ingest and process untrusted data from external legal databases, which creates an indirect prompt injection attack surface.
- Ingestion points: Legal data is retrieved from external APIs including CourtListener, SEC EDGAR, and Congress.gov, as detailed in references/agentic-tech-stack.md.
- Boundary markers: The instructions repeatedly emphasize a mandatory human review gate and draft and verify workflow in SKILL.md and references/agentic-workflows.md to prevent unvetted actions.
- Capability inventory: The proposed agent configurations include capabilities for tool calling, network operations (fetch), and file writing (output_file parameter in references/agentic-tech-stack.md).
- Sanitization: The patterns utilize structured output verification via Zod schemas and Pydantic models to validate machine-readable outputs and mitigate risks.
- [EXTERNAL_DOWNLOADS]: The skill provides instructions for fetching data from well-known legal and government services.
- Fetches case law data from the CourtListener REST API.
- Accesses financial filings via the SEC EDGAR Full Text Search service.
- Retrieves legislative data from the Congress.gov and GovInfo APIs.
- [SAFE]: The skill demonstrates secure development practices, such as using environment variables for API tokens and implementing structured data validation through schemas.
Audit Metadata