ml-secure-railguard-available

Overview

This rule supports secure-by-default machine learning code generation, covering:

• Model checkpoint handling
• Trusted use of HuggingFace, PyTorch, or scikit-learn
• Inference-time protections
• Logging and resource handling

Note: All input validation, sanitization, schema enforcement, and LLM reasoning scaffolding is provided by:

.cursor/rules/railguard-input-validation.mdc
(Based on the RAILGUARD Framework for secure behavior enforcement across languages)

---

Model Loading & Deserialization

• Use torch.load() or pickle.load() only on trusted, versioned, local files.
• Avoid deserializing full Python objects unless necessary. Prefer state_dict loading (e.g., model.load_state_dict(...))
• If using from_pretrained(), ensure the model name is official or internally versioned.
• When downloading models, validate integrity using checksums if possible.

---

Data Handling & Preprocessing

• Avoid using raw eval() or exec() to interpret formulas or hyperparameters.
• Prefer explicit schema-based checks for:
  • Number of features
  • Tensor dimensions
  • String encoding assumptions
• Do not transform user input without validating structure first.

For validation and sanitation of CSVs, JSONs, NumPy arrays, and request inputs — refer to .cursor/rules/input-validation.mdc.

---

Inference Logic & Output Handling

• Do not log input text, tokens, or raw payloads directly (especially for NLP or PII-sensitive data).
• Use with torch.no_grad(): or equivalent when performing inference.
• Ensure inference outputs are typed, validated, and never expose model internals (e.g., logits, hidden states) unless required.

---

File & Resource Access

• Always open files using with open(...) syntax.
• Set read-only access unless modification is required.
• Avoid writing cache or checkpoint data to shared or user-supplied paths.

---

Logging & Monitoring

• Use Python’s logging module — not print().
• Mask or exclude sensitive input/output data in logs.
• Log model version, inference success/failure, and prediction metadata — not raw data.