The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external sources, including customer service tickets (Gorgias, Zendesk), customer reviews (Yotpo, Loox, Trustpilot), and social media mentions. These ingestion points (notably mentioned in customer-service-playbook.md and automation-blueprint-collection.md) lack explicit sanitization protocols or strict boundary markers to separate data from instructions. Given the skill's capability to trigger automated actions such as refunds, email responses, and API calls to platforms like Shopify and Stripe, it is vulnerable to indirect prompt injection where malicious content within a customer ticket or review could manipulate the agent into performing unauthorized actions.
[DYNAMIC_EXECUTION]: The skill documentation, specifically in automation-blueprint-collection.md, outlines architectural plans for executing custom Python and Node.js scripts and utilizing Large Language Models (LLMs) to make decisions based on unstructured data (L4 Automation). This reliance on runtime code assembly and interpretation of external inputs to drive business logic creates a dynamic execution surface that could be exploited if the scripts or LLM inputs are influenced by adversarial data.

afa-ops