The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an inherent attack surface for indirect prompt injection because its primary function involves ingesting and searching external, untrusted data from multiple sources.
Ingestion points: The skill reads local files (CSV, JSON, JSONL, PDF) in scripts/import.py and fetches remote datasets from Hugging Face in scripts/example_data.py. User queries are also interpolated into agent search calls in scripts/ask.py and scripts/query_search.py.
Boundary markers: The scripts do not implement explicit delimiters or instruction-isolation techniques (such as XML tags or "ignore embedded instructions" headers) to prevent the agent from obeying commands hidden within the ingested data.
Capability inventory: The skill possesses capabilities for network operations (reading from Hugging Face and communicating with the Weaviate instance) and local file access. If the agent retrieves malicious instructions from a dataset, it could potentially be tricked into performing unauthorized actions available to its environment.
Sanitization: While scripts/import.py performs basic type validation and casting for structured data fields, there is no NLP-based sanitization for the textual content of the data objects.
[EXTERNAL_DOWNLOADS]: The skill downloads example data from Hugging Face Hub, a well-known service for machine learning assets.
Evidence: scripts/example_data.py uses the datasets library to fetch several repositories, including jamescalam/ai-arxiv2 and AgamiAI/Indian-Income-Tax-Returns. This is a legitimate feature for generating toy data for users.

weaviate