Detection of unauthorised or malicious electromagnetic (EM) transmissions in the wireless spectrum is highly important in both military and commercial systems. In military wireless networks, and particularly in congested EM environments, the detection of unknown radar or communication waveforms can lead to timely identification of potentially adversarial transmissions or intruders in the area. On the other hand, in cognitive radio networks the identification of unauthorised communication waveforms can prevent and mitigate security threats, such as Primary User Emulation (PUE) attacks. However, data of such waveforms are usually of insignificant size to be effectively modelled or even there are no prior data available since they appear for the first time, which makes their timely detection particularly difficult. In this paper, we present a Generative Adversarial Network (GAN) based system which trains on available (presumably friendly) EM signals to detect any previously unseen types of EM waveforms, which can be potentially characterised as unauthorised or malicious. The proposed system is successfully trained and tested on a synthetic dataset comprising different pulsed radar and communication modulated signals impaired with Rician multipath fading, AWGN and random clock offset, resulting in center frequency offset and sampling time drift, and it was shown to successfully detect any previously unseen types of EM waveforms even in low SNR.