The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs the hypogenic package (v0.3.5) from PyPI using the uv package manager. This is a research library developed by ChicagoHAI. The skill documentation provides a thorough security review of the package's dependencies and resource usage patterns.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted text from datasets and literature. It implements a comprehensive defense-in-depth strategy to mitigate potential injections:
Ingestion Points: scripts/audit_dataset.py and scripts/inspect_outputs.py ingest JSON/YAML dataset and result files.
Boundary Markers: Instructions in SKILL.md and references/security.md mandate treating all external text as opaque data and using explicit delimiters in prompts to prevent the agent from following embedded instructions.
Capability Inventory: The bundled local tools are designed without network access or model invocation capabilities. Execution of the actual hypothesis generation software requires separate user approval after auditing.
Sanitization: Scripts utilize schema validation, hash verification (SHA-256), and redaction techniques (e.g., using fingerprints/hashes instead of raw values in reports) to prevent the agent from accidentally executing or echoing malicious instructions embedded in input data.
[DYNAMIC_EXECUTION]: The bundled scripts use a custom UniqueSafeLoader for YAML parsing in scripts/_common.py. This implementation inherits from yaml.SafeLoader and is used to enforce key uniqueness, preventing standard YAML injection attacks while maintaining a safe parsing context.
[COMMAND_EXECUTION]: The skill uses Bash via uv for environment setup and package installation. All command invocations are transparently documented in the SKILL.md file and pinned to specific versions and hashes to ensure reproducibility and integrity.

hypogenic