deep-research
Pass
Audited by Gen Agent Trust Hub on Jul 20, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill identifies a potential indirect prompt injection surface when fetching external data and implements the following mitigations:
- Ingestion points: Uses
web_fetchand optional MCP tools to ingest data from untrusted URLs (SKILL.md). - Boundary markers: Mandates the use of
--- BEGIN UNTRUSTED EXTERNAL CONTENT ---and--- END UNTRUSTED EXTERNAL CONTENT ---to isolate external data from instructions. - Capability inventory: Limited to network fetching (
web_fetch), local database operations (SQL), and text parsing (defuddle). - Sanitization: Includes explicit instructions to the agent to ignore any commands or instructions found within the fetched content.
- [EXTERNAL_DOWNLOADS]: The skill orchestrates the retrieval of information from external websites. It requires user confirmation before performing network operations to prevent issues like rate limiting or unintended data exposure.
- [COMMAND_EXECUTION]: Demonstrates the use of local SQL for research state persistence and provides usage examples for a text-processing CLI (
defuddle). These operations are consistent with the skill's stated purpose of structured data analysis.
Audit Metadata