AI systems often fail unpredictably when inputs change slightly.
Build a suite of tests that automatically vary delay, format, and context of prompts.
AI developers would pay to uncover weaknesses before deployment.
Start with simple timing and rephrasing variations for popular LLMs.
Risk: May require deep model access that providers restrict.