The Agent Skills Directory

[DYNAMIC_EXECUTION]: The 'Calculator' tool provided in the 'Common Patterns' section uses the Python eval() function to process input strings.
Evidence: def calculator(expression: str) -> str: ... return str(eval(expression)) in SKILL.md.
Risk: This allows for arbitrary Python code execution if an attacker provides a crafted string to the agent that gets passed into the tool.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates how to build agents that ingest untrusted data from various external sources without security guardrails.
Ingestion points: Data enters the agent context via tavily.search (web search), vector_store.similarity_search (Redis memory), and contextual.search (RAG integration) as seen in SKILL.md.
Boundary markers: The prompt templates used in the code examples (e.g., web_search_agent and rag_query) lack delimiters or explicit instructions to the model to ignore embedded commands within the retrieved data.
Capability inventory: The agents have the ability to invoke LLMs and execute specialized tools, including the insecure calculator tool.
Sanitization: There is no evidence of input validation, filtering, or sanitization in the provided code snippets to protect against malicious instructions embedded in the external content.
[EXTERNAL_DOWNLOADS]: The skill instructions involve cloning a repository and installing several third-party Python packages.
Evidence: git clone https://github.com/NirDiamant/agents-towards-production.git and multiple pip install commands for libraries like langchain, langgraph, and redis in SKILL.md.
Risk: These dependencies are fetched from standard public registries (GitHub and PyPI) and are consistent with the skill's stated purpose.

agents-towards-production