The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external files (PDF, DOCX, CSV) provided by the user and incorporates this content into generated reports, creating a surface for indirect prompt injection.
Ingestion points: Reads content from user-provided files within the sources/ directory using utilities like pdftotext, textutil, and LibreOffice.
Boundary markers: Uses {{placeholder}} syntax to mark unverified data and data-sources attributes to link report sections to their original source IDs.
Capability inventory: The skill can execute various subprocesses including google-chrome for PDF generation, osascript for Pages conversion, and document processing tools like pdftoppm and pdftotext.
Sanitization: Implements a custom HTML parser and linter that restricts the report.html file to a strict whitelist of allowed elements and attributes, specifically prohibiting <script> and style tags.
[EXTERNAL_DOWNLOADS]: The dependency check script automatically installs the docx library from the public NPM registry if it is not already present.
Evidence: The check_deps.mjs script executes npm ci to install dependencies defined in scripts/package.json.
[COMMAND_EXECUTION]: The skill relies on several external CLI tools for its core functionality, which are executed as subprocesses.
Evidence: Uses spawnSync to run google-chrome, pdftotext, pdftoppm, pdffonts, soffice, and osascript with arguments derived from local file paths.
[PRIVILEGE_ESCALATION]: The skill instructions direct the agent to request user permission to run commands outside the agent's restricted sandbox environment if the sandbox blocks the execution of required tools.
Evidence: SKILL.md contains specific instructions for the agent to handle sandbox restrictions by requesting permission from the user to run commands like Chrome or npm outside the seatbelt sandbox environment.

global-report