The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process unstructured data from external registries such as HuggingFace, arXiv, and various vendor documentation sites, which could potentially contain malicious instructions.
Ingestion points: The skill fetches content from URLs defined in references/source-registry.yaml, including model cards and evaluation tables.
Boundary markers: The instructions in SKILL.md explicitly state that external content must be treated as data and warn against following embedded instructions. The schema schemas/ml-task-vocabulary-map.schema.json requires a sanitization field to record the posture taken during retrieval.
Capability inventory: The skill's primary capabilities are network retrieval and file writing to generate survey reports.
Sanitization: The workflow enforces a sanitization protocol where potentially harmful content (e.g., agent-directed notes) is stripped before the data is processed, with the outcome recorded in the final artifact.
[COMMAND_EXECUTION]: The test suite scripts/test_validate_ml_prior_art.py utilizes shell commands to verify the functionality of the validator script.
Evidence: subprocess.run() is used in the TestCLI class to execute the validate_ml_prior_art.py script and check its return codes and output. This is a standard and benign use for testing command-line tools.
[EXTERNAL_DOWNLOADS]: The skill references and fetches data from numerous well-known machine learning registries and academic sources.
Evidence: The references/source-registry.yaml file configures access to sources including HuggingFace Hub, arXiv, OpenML, Zenodo, and official documentation for OpenAI, Google AI, AWS, and Microsoft Azure. These are established services used for their intended purpose in the survey workflow.

ml-prior-art-survey