The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses subprocess.run in scripts/container_meta.py, scripts/image_meta.py, and scripts/rewrite_text.py to execute other scripts and system binaries like exiftool, qpdf, and c2patool. It dynamically determines paths to interpreters (e.g., locating virtual environments) and handles subprocess execution for image and PDF processing, which constitutes dynamic execution of local code and system tools.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted files (PDF, HTML, Markdown, Images) and passes their text content to LLMs for rewriting in scripts/rewrite_text.py. It lacks explicit boundary markers or sanitization within the prompt templates used for rewriting, creating an attack surface where malicious instructions in a processed file could influence the agent's behavior. The capability inventory for this surface includes subprocess execution, network operations, and file system writes.
[COMMAND_EXECUTION]: The skill executes external command-line tools via subprocess.run to perform its primary functions, such as qpdf for PDF structural rewrites and exiftool for metadata stripping. These calls are protected using path-guarding logic (safe_arg) to prevent flag injection.
[DATA_EXFILTRATION]: scripts/rewrite_text.py uses urllib.request to send text content to user-configured AI service endpoints. While the implementation defaults to loopback addresses, includes protections against HTTP redirects to prevent credential leakage, and requires an explicit override for remote hosts, it represents a network interface for data processed by the skill.

remove-claude-marks