The Agent Skills Directory

[DYNAMIC_EXECUTION]: The references/langchain-patterns.md file includes a code pattern for a calculator_tool that utilizes the eval() function. This function executes arbitrary Python code from strings, creating a severe vulnerability to remote code execution (RCE) if the agent processes untrusted input through this tool.
[CREDENTIALS_UNSAFE]: The documentation in references/data-versioning.md contains hardcoded sample AWS access keys (AKIAIOSFOLQUICKSTART and a sample secret key). While these are standard documentation placeholders, they trigger high-severity security detections and represent poor secret management practice.
[COMMAND_EXECUTION]: The implementation in examples/feast-features/setup_features.py utilizes os.system() to execute shell commands for initializing and applying feature store configurations. Spawning shell processes via os.system() is a risky pattern that can lead to command injection if inputs are not strictly controlled.
[INDIRECT_PROMPT_INJECTION]: The skill architecture is designed to build RAG pipelines that ingest external documents (PDF, Markdown, HTML) and inject them into LLM prompts without implementing content sanitization or robust instruction boundaries.
Ingestion points: Document loaders in examples/dagster-pipelines/embedding_pipeline.py (loading from a source directory) and examples/langchain-rag/main.py (using TextLoader).
Boundary markers: Missing or inadequate; prompt templates use standard delimiters but lack explicit instructions for the model to ignore embedded malicious commands within the retrieved context.
Capability inventory: The skill uses vector database writes, LLM generation, and file system operations.
Sanitization: No evidence of content filtering or sanitization was found in the provided implementation examples.

ai-data-engineering