The Agent Skills Directory

[DYNAMIC_EXECUTION]: The code example provided in the 'Eval-Gated CI/CD' section of SKILL.md uses the eval() function to process text generated by an AI model (gpt-5.5). Since model outputs are influenced by external inputs and can be manipulated via prompt injection, using eval() on this content allows for arbitrary code execution on the system running the evaluation logic.
[INDIRECT_PROMPT_INJECTION]: The guide describes architectures for RAG and agentic systems that ingest untrusted user and document data.
Ingestion points: Data ingestion occurs in the retrieval logic and agent task loops described in SKILL.md where user queries and corpus text are processed.
Boundary markers: While the guide mentions action gates and sandboxes in specific examples, these markers are not uniformly integrated into all provided design patterns, potentially allowing embedded instructions to be followed.
Capability inventory: The patterns include capabilities for performing network operations and tool execution through various AI frameworks and protocols like MCP.
Sanitization: The provided code snippets demonstrate a lack of strict output validation or sanitization before passing model-generated content to execution functions like eval().

ai-system-design-guide