ml-secure-no-railguard-available

R: Risk First

• The goal is to reduce risk exposure in machine learning pipelines by securing data ingestion, model checkpoint loading, dependency handling, and inference logic.
• AI-generated ML code should not trust unknown data sources or model files without validation.
• The LLM must ensure safety against insecure serialization, poisoning, and unsafe dynamic behavior.

A: Attached Constraints

• Never use pickle.load() or torch.load() on files from untrusted sources.
• Never use eval() or exec() to dynamically interpret model code, formulas, or input.
• Never suppress exceptions silently (try/except: pass).
• Avoid using untyped data transformations or outputs.
• Do not log raw inputs from end users (can include PII).

I: Interpretative Framing

• Treat all input data (CSV, JSON, NumPy, HuggingFace datasets) as potentially malformed or poisoned unless explicitly validated.
• If loading a model checkpoint, assume the file may have been tampered with.
• When generating inference code, assume it may be deployed in production with untrusted input.

L: Local Defaults

• Use joblib or torch.load() only on trusted, versioned model paths
• Prefer torch.save(model.state_dict()) for safe model export; avoid full object serialization
• Use Pydantic or Marshmallow schemas for preprocessing configs and inference input validation
• Use logging for monitoring; avoid print() and never log raw request.body
• Default to strict file permissions (r, no rb+)

G: Generative Path Checks

1. When generating model loading logic:
   • Confirm source is trusted or version-controlled
   • Avoid deserializing entire objects unless safe
   • Use checksum or hash verification if relevant
2. When preprocessing data:
   • Validate data structure (rows, types, shape)
   • Use try/except with logging for failed transforms
3. When handling input for inference:
   • Validate schema
   • Normalize securely
   • Avoid leaking model internals in output

U: Uncertainty Disclosure

• If unsure about input format, file source, or serialization method, generate a comment: “Verify this file path is trusted before deserializing model.”
• If unsure about preprocessing correctness, generate: “Review schema/shape assumptions before transforming user input.”