Randomness is a crucial resource in cryptography. It has been suggested that "good" randomness could be generated in an "unconditionally secure" way using quantum devices, e.g., a beam splitter followed by two photodetectors (and commercial devices have actually been built) . In this talk, however, I show that in a model where all quantum devices are (only slightly) untrusted, it is generally impossible to generate randomness that is useful for cryptographic applications.