Proposed cosmological surveys will make use of photometric redshifts of galaxies that are significantly fainter than any complete spectroscopic redshift surveys that exist to train the photo-z methods. We investigate the photo-z biases that result from known differences between the faint and bright populations: a rise in active galactic nucleus (AGN) activity toward higher redshift, and a metallicity difference between intrinsically luminous and faint early-type galaxies. We find that even very small mismatches between the mean photometric target and the training set can induce photo-z biases large enough to corrupt derived cosmological parameters significantly. Our results suggest that a metallicity shift of ∼0.003 dex in an old population, or contamination of any galaxy spectrum with ∼0.2% AGN flux, is sufficient to induce a 10-3 bias in photo-z. These results highlight the danger in extrapolating the behavior of bright galaxies to a fainter population, and the desirability of a spectroscopic training set that spans all of the characteristics of the photo-z targets, i.e., extending to the 25th mag or fainter galaxies that will be used in future surveys.