The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses prompt templates that interpolate untrusted inputs (user text and retrieved context) directly into strings without using robust delimiters or security instructions.
Ingestion points: Found in references/llm_integration_guide.md (e.g., {user_input}, {product_context}) and references/rag_system_architecture.md (e.g., {context}, {question}).
Boundary markers: Absent; the templates rely on standard headers like 'Context:' and 'Question:', which are easily bypassed by adversarial content.
Capability inventory: The skill includes scripts for model deployment and orchestration (scripts/model_deployment_pipeline.py, scripts/rag_system_builder.py) that could be targeted if an agent follows malicious instructions in the retrieved context.
Sanitization: No sanitization, validation, or escaping logic is applied to the input variables before they are interpolated into the prompt strings.
[COMMAND_EXECUTION]: The skill includes local Python scripts designed to automate MLOps workflows which are intended to be executed from the CLI.
Evidence: SKILL.md provides commands to run scripts/model_deployment_pipeline.py, scripts/rag_system_builder.py, and scripts/ml_monitoring_suite.py with various arguments.

senior-ml-engineer