The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads research datasets from the official 'ChicagoHAI' organization (Human-AI Interaction Lab) on GitHub and installs the 'hypogenic' package from PyPI. These sources are consistent with the skill's purpose for scientific discovery and research.- [COMMAND_EXECUTION]: Instructions guide the user or agent to execute shell commands for environment setup and data management, including 'uv pip install' for dependencies, 'git clone' for datasets, and bash scripts for PDF preprocessing services like GROBID. These are routine tasks for scientific software but involve executing code sourced from external repositories.- [PROMPT_INJECTION]: Indirect Prompt Injection Surface. The core functionality involves ingesting text from external JSON datasets and interpolating it into LLM prompts to generate and test hypotheses. This presents an attack surface for indirect prompt injection. 1. Ingestion points: Research data is read from JSON files specified in the task configuration (SKILL.md). 2. Boundary markers: Analysis of the 'config_template.yaml' reveals that data placeholders such as {sample_text} and {data_samples} are used without explicit isolation markers or delimiters in the prompt templates. 3. Capability inventory: The skill has the capability to execute shell commands and communicate with LLM provider APIs. 4. Sanitization: No evidence of data sanitization or specific instruction-filtering logic is present in the static configuration.

alterlab-hypogenic