source-evidence-dataset-development
Pass
Audited by Gen Agent Trust Hub on Sep 1, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill uses the
npx skillscommand to resolve and run a specific extension (tiangong-kb-sci-search) from the vendor's repository athttps://github.com/tiangong-ai/skills. This is an intended mechanism for expanding research capabilities within the vendor's framework. - [INDIRECT_PROMPT_INJECTION]: The skill ingests data from multiple untrusted sources, which could contain adversarial instructions.
- Ingestion points: Processes documents (PDF, Word), web pages, URLs, APIs, and academic literature.
- Boundary markers: The instructions require the agent to capture limitations and conflicts separately from field evidence and use a structured evidence dossier for all authored fields.
- Capability inventory: Includes file reading, workspace management, CLI tool execution, and remote skill resolution.
- Sanitization: All authored data must pass through a CLI-based validation and QA curation queue before commitment.
- [COMMAND_EXECUTION]: The instructions direct the agent to execute the
tiangong-lcaCLI tool for authentication health checks and data validation tasks. The skill explicitly prohibits the agent from requesting or handling user credentials directly, delegating authentication to a trusted terminal session.
Audit Metadata