Hyperparameters are chosen by search over a validation set, which makes them exactly as dangerous as strategy parameters. Every combination tried is another test, and the best result across a grid of 200 is an optimistic estimate by construction.
This is why the train-validation-test-split has three parts. Train fits the coefficients, validation chooses the hyperparameters, and test is touched once at the end. Collapsing validation and test is the most common way machine-learning results in trading become fiction.
Tuning also interacts with purged-cross-validation: the purging and embargo must be applied inside every fold of the hyperparameter search, not just once at the end, or the tuning itself leaks.
Related: parameter, regularisation, train-validation-test-split, bayesian-optimisation