The quantum hat, not the head

When a "quantum" model beats a classical one, the right question is not who won but what part of the model actually carried the score. A new audit of quantum machine learning for catching network intruders forces that question, and the answer is mostly unflattering to the label. Vendors are selling "quantum-powered" intrusion detection while ordinary random forests and gradient-boosted trees, properly tuned, dominate the headline numbers across four standard datasets. The "quantum" turns out to be mostly the hat, not the head.

The reusable category is attribution. A claimed advantage survives only when a matched classical control cannot reproduce it under the same feature budget, leakage rules, and false-discovery correction. Most quantum edges in the new audit do not survive that test. They dissolve into preprocessing, regularization, and an unbalanced metric choice.

Two narrow effects do survive. A quantum-kernel model out-ranks a directly comparable random-feature kernel on calibrated metrics. And a small four-qubit hybrid beats the best classical model at the strictest low-false-alarm setting on one distribution-shifted dataset: the one audited residual quantum effect the authors flag in the preprint.

The mechanism generalizes. Any "quantum" or "AI" advantage in tabular security data should be priced against an honest classical control, not a strawman, and tested at the operating point the buyer actually cares about. Security teams shopping for signal rather than labels win. Marketing that depends on the quantum halo loses. The buyers' question shifts from "is it quantum?" to "which kernel, at which false-positive budget, on which distribution-shifted data, with what multiple-comparison correction?"