This is just speculation on your end, but even it's true, a model choosing to hack against instructions simply because it knows how to hack is by definition misalignment.