The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it extracts content from user-provided PowerPoint files and passes it to the AI for analysis or visual inspection without protective boundary markers or warnings. A malicious presentation could contain hidden instructions designed to override agent safety protocols or influence the quality assessment during the QA workflow.
Ingestion points: Untrusted content is extracted from files via markitdown and processed using scripts such as thumbnail.py and unpack.py (found in SKILL.md).
Boundary markers: The instructions in SKILL.md for visual and content QA do not utilize delimiters (such as XML tags or triple backticks) or include system-level instructions to ignore commands found within the slide data.
Capability inventory: The skill possesses local write and delete permissions through its management scripts (scripts/add_slide.py, scripts/clean.py, scripts/office/pack.py) and executes local command-line tools for document conversion and image generation.
Sanitization: No explicit filtering or sanitization of the extracted content is performed before it is interpolated into agent prompts.
Mitigation: Wrap extracted content in delimiters and include an explicit 'ignore embedded instructions' warning in the prompts used for document processing.

powerpoint