Basically what happens is that Windows switches to the secure desktop, makes it the current one, so input is now associated with it.
The old desktop remains where it was: all the HWNDs on the desktop are still there, and any thread attached to that desktop can still access those HWNDs, get their location, and so on. You can still send messages to windows on this desktop, so long as the thread sending the message is also on that desktop.
However, since the desktop is now inactive, it cannot receive input. GetForegroundWindow will return NULL (IIRC), and you can't use SendInput any longer, since input now belongs to [a thread on] a different desktop; no controls on that inactive desktop can receive focus.
Note that sending keypress messages to a control that doesn't have focus can sometimes cause unexpected behavior, since the app or control generally never expects to receive keyboard input without getting the focus first. (This can be problematic for controls that set up some sort of input context in WM_SETFOCUS and clear it up in WM_KILLFOCUS, for example.)
In short, the UI is still there: you can do certain queries against it, but you can no longer automate it as you could on a regular desktop by sending input, and some other functions that relate to focus or input may fail.
I'm not super familiar with AutoHotKey, but the name and description of functionality suggests that it's heavily reliant on the underlying Win32 SendInput API. This won't work at all for keyboard input when a desktop is inactive.
For a reasonable overview of how desktops work and how they relate to winstations, the locked desktop, and so on, check out the Desktop article on MSDN.
One issue that I've run into in the past with desktops and automation is: how to I leave a long-running test that's using some form of user input automation (mouse, keyboard simulation), but still lock my PC so that someone can't just walk by and interfere with it. Once you lock the PC, the desktop is inactive, and so the automation stops working. A similar issue happens if the screensaver kicks in: the desktop switches, and the automation fails.
One solution is to use two PCs: let's call them Main and Test: from Main, open a remote terminal services client onto the Test machine, and then run the automated test on the test machine, but from a terminal services client window on the Main machine. Now the cool part: you can minimize that TSC window, or even lock the Main machine (or let the screensaver kick in), and that virtual session will continue working, thinking that it is still active - it's just that nobody is paying it any attention. This is one way to create a "connected" session with an active desktop, but one that no-one can interfere with, because it's protected behind the locked desktop of the Main machine.
