AI risk & limits

Where you must put a human in the loop.

The idea

AI is a massive productivity booster, but it poses three major risks: Data Exfiltration (pasting API keys into public models), Copyright/Licensing (generating code derived from GPL), and Silent Failures (hallucinating logic in a medical or financial app).

Before handing a task to an AI, you must run it through a Risk Gate to determine if it can be done autonomously, if it requires strict human review, or if AI is strictly forbidden.

Does the context window contain:

API Keys, Passwords, or PII (Customer Data)? Mission Critical Logic (Payments, Security)? Standard Boilerplate (CSS, Forms, Config)?

Pending Assessment

Select the parameters above.

Evaluate the risk profile before sending code to the model.

How it works (Data & Logic Risks)

# Risk 1: The Training Data Leak
# Public LLMs may retain your prompts to train future models.
# NEVER paste AWS keys, user emails, or proprietary algorithms
# into a public chat window.

# Risk 2: The "Looks Right" Trap
# AI code is highly plausible. If you use it for cryptography
# or payment calculation, a subtle hallucination will cause
# a catastrophic breach. Human review is mandatory.

# Risk 3: IP Contamination
# AI might spit out exactly copied GPL code. If you paste that
# into your proprietary codebase, you might violate licenses.