The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_ai_alignment.py utilizes subprocess.run to execute external Python scripts for rendering PDF and Word documents. These scripts are located within a dependency skill (research-literature-review). While the skill uses list-based argument passing to mitigate command injection, the paths to these scripts are resolved dynamically at runtime.
[DYNAMIC_EXECUTION]: The skill performs dynamic path resolution in scripts/runtime_utils.py to locate dependency skills and their associated scripts. This behavior, combined with the execution of discovered scripts via subprocesses, represents a dynamic execution pattern common in modular agent skills.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from BibTeX files and PDF documents, extracting abstracts and text excerpts to provide context for the AI agent's analysis. This represents an indirect prompt injection surface where malicious instructions embedded in research papers could influence the agent's behavior. 1. Ingestion points: Reads .bib and .pdf files in scripts/bib_utils.py. 2. Boundary markers: The skill generates a structured JSON containing the data for the AI, but relies on the AI to distinguish between the content and instructions. 3. Capability inventory: File write operations and subprocess execution via scripts/run_ai_alignment.py. 4. Sanitization: Minimal sanitization is performed on the extracted text before it is passed to the AI agent.

research-citation-check