The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external file formats such as PDF, DOCX, and XLSX which are ingestion points for untrusted data. Since the output is intended for consumption by an AI agent, this creates a surface for indirect prompt injection where malicious instructions embedded in documents could influence agent behavior. While scripts like batch_convert.py and convert_literature.py use Markdown headers as boundary markers, no explicit sanitization is performed on the converted content. The risk is compounded by the skill's Read, Write, Edit, and Bash permissions.
[PERSISTENCE]: The INSTALLATION_GUIDE.md and OPENROUTER_INTEGRATION.md files instruct users to append environment variables and API keys to shell profiles like ~/.bashrc and ~/.zshrc. This represents a persistence mechanism, although documented here as a legitimate user setup step for credential management.
[PROMPT_INJECTION]: The SKILL.md file contains behavioral directives instructing the agent to always consider adding scientific diagrams and states that schematics should be generated by default. These instructions attempt to override the agent's default task logic by mandating additional sub-tasks.

markitdown