from-pdf-skill-builder
Pass
Audited by Gen Agent Trust Hub on Jun 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes local shell commands (pdftoppm, pdftotext, sed, grep, mv) to process PDF files and organize the resulting skill structure. These operations are performed within the local skill environment and align with the stated purpose of document extraction and skill assembly.\n- [PROMPT_INJECTION]: As the skill involves subagents processing untrusted data from PDFs via vision, it presents an indirect prompt injection surface. Malicious content within a source PDF could attempt to influence the subagent's transcription task. This risk is managed by the provided templates which enforce strict transcription rules ("Faithfulness rules") and clearly defined task boundaries.\n
- Ingestion points: User-supplied PDF documents processed by subagents as described in SKILL.md.\n
- Boundary markers: The vision-extraction template (templates/subagent-vision-prompt.md) defines explicit "Inputs", "Output", and "Process" sections to separate instructions from the data being processed.\n
- Capability inventory: Subagents are granted access to the Bash and Read tools to handle image rendering and text transcription tasks.\n
- Sanitization: The skill relies on "Faithfulness rules" instructing the model to transcribe content verbatim and preserve original context, which serves as a control against the execution of instructions embedded in the document text.
Audit Metadata