book-sft-pipeline
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a data processing pipeline that extracts text from external book files (ePub) and interpolates this content into a prompt for an LLM to generate training instructions. This creates a vulnerability surface where a malicious book could theoretically attempt to influence the instruction generation phase.
- Ingestion points: ePub text extraction in
SKILL.md(Phase 1) and segmentation logic inscripts/pipeline_example.py. - Boundary markers: Text chunks are interpolated into the
INSTRUCTION_PROMPTusing the{text}placeholder inSKILL.mdwithout additional sanitization or isolation markers. - Capability inventory: The skill utilizes an LLM (e.g., Gemini Flash) for generating scene descriptions and uses the Tinker API to perform LoRA fine-tuning on base models.
- Sanitization: No explicit content filtering or escaping is applied to the extracted text before it is processed by the instruction-generation LLM.
Audit Metadata