We present the first security analysis of conference key agreement (CKA) in the most adversarial model of device independence (DI). Our protocol can be implemented by any experimental setup that is capable of performing Bell tests (specifically, the Mermin-Ardehali-Belinskii-Klyshko (MABK) inequality), and security can in principle be obtained for any violation of the MABK inequality that detects genuine multipartite entanglement among the N parties involved in the protocol. As our main tool, we derive a direct physical connection between the N -partite MABK inequality and the CHSH inequality, showing that certain violations of the MABK inequality correspond to a violation of the CHSH inequality between one of the parties and the other N − 1. We compare the asymptotic key rate for DICKA to the case where the parties use N − 1 DIQKD protocols in order to generate a common key. We show that for some regime of noise the DICKA protocol leads to better rates.Quantum communication allows cryptographic security that is provably impossible to obtain using any classical means. Probably the most famous example of a quantum advantage is quantum key distribution (QKD) [1,2], which allows two parties, Alice and Bob, to exchange an encryption key whose security is guaranteed even if the adversary has an arbitrarily powerful quantum computer. What's more, properties of entanglement lead to the remarkable feature that security is sometimes possible even if the quantum devices used to execute the protocol are largely untrusted. Specifically, the notion of device independent (DI) security [3-5] models quantum devices as black boxes in which we may only choose measurement settings and observe measurement outcomes. Yet, the quantum state and measurements employed by such boxes are unknown, and may even be prepared arbitrarily by the adversary.Significant efforts have been undertaken to establish the security of device independent QKD [5][6][7][8][9][10][11], leading to ever more sophisticated security proofs. Initial proofs assumed a simple model in which the devices act independently and identically (i.i.d.) in each round of the protocol. This significantly simplifies the security analysis since the underlying properties of the devices may first be estimated by gaining statistical confidence from the observation of the measurement outcomes in the tested rounds. The main challenge overcome by the more recent security proofs [8][9][10][11] was to establish security even if the devices behave arbitrarily from one round to the next, including having an arbitrary memory of the past that they might use to thwart the efforts of Alice and Bob. Assuming that the devices carry at least some memory of past interactions is an extremely realistic assumption due to technical limitations, even if Alice and Bob prepare their own trusted, but imperfect, devices, highlighting the extreme importance of such analyses for the implementation of device independent QKD. In contrast, relatively little is known about device independence outside the realm ...