The Agent Skills Directory

[DYNAMIC_EXECUTION]: The references/agentic-workflows.md file contains a code example for a calculation tool that uses the JavaScript eval() function on input provided to the model. This pattern is inherently unsafe as it allows the model to execute arbitrary code if manipulated by a malicious user prompt.
[INDIRECT_PROMPT_INJECTION]: The skill implements RAG and agent workflows that pull content from external sources (vector databases, web search) and interpolate it directly into prompts, creating an attack surface for indirect instructions.
Ingestion points: Vector database results in templates/rag-pipeline-template.ts and web search results in templates/agent-workflow-template.ts.
Boundary markers: The templates use natural language instructions (e.g., 'Answer ONLY based on context') but lack distinct structural delimiters or escaping for the external data.
Capability inventory: The system includes tools for email sending, database access, and web search, providing high-value targets for an injection attack.
Sanitization: There is no evidence of filtering or sanitizing retrieved content before it enters the model's context.
[EXTERNAL_DOWNLOADS]: The references/ollama-local-inference.md file provides instructions to download and execute an installation script from Ollama's official domain (https://ollama.com/install.sh). This is a standard installation procedure for a well-known service.

ai-native-development