The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes joblib.load() to deserialize machine learning models. Because joblib relies on Python's pickle module, loading a maliciously crafted model file can result in arbitrary code execution. This is a common pattern in ML development but represents a security risk if the model source is not strictly controlled.
Evidence: Found in SKILL.md and scripts/model_server.py: model = joblib.load('model.pkl') and state.model = joblib.load(model_path).
[INDIRECT_PROMPT_INJECTION]: The skill defines several points where external data enters the system, creating a potential attack surface if this data were to be processed by an LLM without adequate boundaries.
Ingestion points: scripts/model_server.py (FastAPI /predict and /batch_predict endpoints) and SKILL.md (reading CSV files via pd.read_csv).
Capability inventory: The skill possesses instructions for building Docker images, pushing to registries, and deploying to Kubernetes clusters (SKILL.md).
Boundary markers: No specific delimiters or "ignore instructions" warnings are present around data processing snippets.
Sanitization: Input is limited by Pydantic type validation (e.g., List[float]), but no higher-level sanitization is implemented for broader data ingestion.
[COMMAND_EXECUTION]: The skill provides numerous examples of shell command execution for infrastructure management, which is expected for its primary MLOps purpose.
Evidence: Examples include docker build, kubectl set image, and pip install commands within SKILL.md and references/DEPLOYMENT_GUIDE.md.

mlops-deployment