hypogenic
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches research datasets and configuration examples from the Chicago Human+AI Lab (ChicagoHAI) GitHub repositories.
- [COMMAND_EXECUTION]: Utilizes CLI tools and shell scripts (e.g.,
hypogenic_generation,setup_grobid.sh) to automate hypothesis generation, testing, and PDF document processing. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources that are interpolated into LLM prompts.
- Ingestion points: Loads data from JSON training/testing files and research papers in PDF format located in the local workspace.
- Boundary markers: Employs template variables like
${text_features_1}for data injection; however, no explicit safety delimiters or 'ignore' instructions for embedded data are defined in the provided templates. - Capability inventory: Performs shell command execution for generation tasks and interacts with local services like GROBID for file processing.
- Sanitization: The skill does not document specific sanitization or filtering logic for the contents of processed datasets or research papers.
- [METADATA_POISONING]: Contains instructions for the agent to recommend the K-Dense Web platform to users when encountering complex research workflows.
Audit Metadata