Bug Detection Blind Spots in AI Coding Harnesses (GStack and Beyond)

AI coding assistants surprisingly struggle less with a bug's complexity and more with situations where crucial information is missing. In a series of 28 debugging experiments, AI models consistently fixed difficult bugs involving deep internals and numerical edge cases when all necessary context was available in the code. However, every model failed on a seemingly simple bug that required knowledge of an undocumented API contract, producing incorrect fixes that passed all tests but silently corrupted user data. This highlights a critical failure mode where an AI cannot recognize its own knowledge gaps, confidently submitting flawed solutions that appear correct and can easily slip into production.