The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill design involves fetching data from external, untrusted web sources and passing it to an LLM for analysis, which creates a potential for indirect prompt injection. * Ingestion points: Data is collected in scraper/sources/source_name.py from external URLs or APIs and processed in ai/pipeline.py. * Boundary markers: The _build_prompt function in ai/pipeline.py uses markdown headers to separate data from instructions but lacks explicit delimiters or instructions to ignore potentially malicious content embedded in the scraped data. * Capability inventory: The resulting agent has network access for scraping and communicating with the Gemini and Notion APIs, file system write access for local feedback persistence, and scheduled execution via GitHub Actions. * Sanitization: Items are serialized via json.dumps before being interpolated into the prompt, and the architecture includes a filters.py module for rule-based keyword filtering.

data-scraper-agent