ai-rag-pipeline
Pass
Audited by Gen Agent Trust Hub on Aug 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill demonstrates the use of the
beltCLI tool (which is scoped in the skill's allowed-tools configuration) to execute search and analysis applications on the inference.sh platform.- [EXTERNAL_DOWNLOADS]: Provides links to installation guides and documentation hosted on the officialinference.shdomain and its associated GitHub repositories.- [PROMPT_INJECTION]: As a research tool, the skill documents patterns for ingesting untrusted external data into LLM prompts, which is the standard mechanism for RAG systems. - Ingestion points: The examples in
SKILL.mdshow external search data being captured into shell variables such as$SEARCH_RESULT,$CONTENT, and$EVIDENCE. - Boundary markers: The skill uses textual headers (e.g., 'Search Results:', '== Statistics ==') to separate retrieved context from instructions, but does not implement programmatic delimiters or safety warnings for the model.
- Capability inventory: Aggregated data is sent to external large language models (Claude, GPT-4, etc.) via the
belt app runcommand. - Sanitization: External content is interpolated directly into model prompts without evidence of automated sanitization or filtering.
Audit Metadata