The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process unstructured data from external sources, creating a potential surface for indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context through product_description in the Product Classification class, invoice_text in the Invoice Extractor class, news articles in the Supplier Risk Analyzer, and user_input in the Chatbot implementation.
Boundary markers: There are no boundary markers or instructions to the models to disregard embedded commands within the input text (e.g., in SKILL.md).
Capability inventory: The implementation currently includes data classification, named entity extraction, and response generation; it does not explicitly demonstrate dangerous capabilities like file writing or network exfiltration of sensitive data within the provided snippets.
Sanitization: The skill lacks input sanitization or validation logic to filter potentially malicious instructions embedded in the processed text.

nlp-supply-chain