Where AI coding assistants get it wrong.

Assistants are trained on code that is predominantly correct at the functional level — it does what it was asked to do. But most defects that matter are not visible in the happy path. They live in error paths, in concurrent requests, at the boundary between layers, and in the assumptions baked into an implementation choice nobody stated.

So a model asked to “add authentication” or “paginate this endpoint” produces something that passes a reasonable functional test — and rarely considers what happens on the second concurrent request, the millionth row, or the request that bypasses the route entirely.

We call the result implementation gaps: controls and code paths that work under expected conditions and fail under adversarial ones. They are the hardest kind to find, because they do not produce obvious bugs. The tests pass. The demo works. The code looks reasonable.