I've designed some processors, albeit not at the complexity of Intel, AMD, or IBM. However, in theory it's definitely possible via any number of mechanisms. This could be (provably) undetectable by software state, though perhaps measuring e.g. deviance from average clock cycles of certain instructions to determine whether something unusual was happening. For example, this is almost certainly similar to what would be used: http://www.intel.com/content/www/us/en/architecture-and-tech...
To me, this is the boring part—I'm much more interested in a) what they would collect, b) how they would identify it from the processor (if they didn't load software into memory), and c) how they expect to retrieve it. If they even attempted to use IP to communicate they would be caught immediately.