Hallucination is what happens when the spec isn't there, and niche APIs are exactly where the docs were thin before any model touched them. The pattern-based parts are safe because there's a corpus to interpolate from; the dangerous parts are the endpoints nobody wrote down, where the model is guessing and confident about it. Giving the agent a sandbox to poke at doesn't fix that if the sandbox is built from the same absent spec, because you've just automated the guessing.