The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided PDF files to apply visual modifications. This creates an attack surface where adversarial content embedded in the PDFs could attempt to influence the agent's behavior during the inspection workflow defined in the skill instructions.
Ingestion points: The skill ingests untrusted data via the input.pdf and sig_pdf arguments in scripts/make_scanned.py. The SKILL.md file instructs the agent to "Examine the source signature PDF" and "Examine the target PDF" to identify locations for modification.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are provided in the prompt instructions to isolate the content of the processed PDFs from the agent's control logic.
Capability inventory: The skill possesses the capability to render PDFs to images (using pdf2image and poppler), perform image manipulation via pillow and numpy, and write new files to the filesystem via img2pdf.
Sanitization: The skill does not perform any sanitization, filtering, or validation of the text or metadata content within the input PDFs before the agent is prompted to examine them.
[EXTERNAL_DOWNLOADS]: The skill relies on several external dependencies and system utilities.
Evidence: SKILL.md instructions include the installation of Python packages pillow, pdf2image, img2pdf, and numpy. It also directs users to download poppler binaries from a GitHub repository (github.com/osber/poppler-windows). These are standard requirements for the skill's stated purpose.
[COMMAND_EXECUTION]: The script modifies the PIL.Image.MAX_IMAGE_PIXELS setting to 300,000,000 in scripts/make_scanned.py to handle high-resolution PDF rendering, which technically bypasses a default security limit intended to prevent decompression bomb (DoS) attacks, though this is a common requirement for high-quality document processing.

pdf-look-scanned