A team of researchers from the US and Europe has demonstrated a "hardware trojan" attack on Intel's Ivy Bridge processor. This paper (PDF) describes the exploit, which changes the dopant polarity of individual transistors to weaken the chip's random number generator. The researchers were able to reduce the random number generator's entropy from 128 to 32 bits, making cryptographic keys much more predictable. They claim the exploit is stealthy enough to pass not only the CPU's built-in self-test, but also the National Institute of Standards and Technology's tests for random number generators.
Inserting the trojan involved altering the dopant masks of "only a few" transistors. Ivy Bridge has about 1.4 billion transistors, making the small change difficult to detect. According to the researchers, the "sub-transistor" trojan can't be exposed by optical reverse engineering because the chip's circuitry remains the same.
Ars Technica has a good summary of the paper, including some commentary from one of its authors. That researcher points out that hardware trojans haven't been found in the wild. However, the proof-of-concept attack illustrates that existing chips are vulnerable to hardware exploits that may be impossible to detect.