The Agent Skills Directory

Dynamic Execution (LOW): The skill uses joblib.load to deserialize model files. Evidence: model = joblib.load("model.pkl") in SKILL.md. Context: While standard in machine learning workflows, Joblib (based on Pickle) can execute arbitrary code during deserialization. This finding is downgraded from MEDIUM as it is central to the primary purpose of the skill.
Indirect Prompt Injection (LOW): The skill implements API endpoints that ingest untrusted external data. Ingestion points: PredictRequest features in SKILL.md. Boundary markers: Absent. Capability inventory: Use of ProcessPoolExecutor for parallel task execution. Sanitization: Basic type validation via Pydantic; no semantic validation or sanitization of input features.

ml-engineer