The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its ingestion of untrusted external content.
Ingestion points: The skill accepts external data through materialText, materialFile, and materialUrl parameters as defined in references/form-schema.json.
Boundary markers: There are no explicit delimiters or system instructions provided to the model to ignore potential malicious commands that might be hidden within the customer materials or linked documents.
Capability inventory: The skill utilizes a Python runner (scripts/run.py) to transmit this data to an external processing API (ai-skills.ai).
Sanitization: No input validation or sanitization is present in the skill's scripts to filter or escape instructions before processing.
[EXTERNAL_DOWNLOADS]: The skill includes a Python script that facilitates network communication with external services.
Evidence: The script scripts/run.py uses urllib.request to send user data and parameters to the vendor's API at https://ai-skills.ai for processing.
Context: These network operations are directed at the vendor's infrastructure to provide the skill's primary functionality and use environment variables for secure API key management.

ai-sales-enablement