THE QUANTUM INTELLIGENCER

An eavesdropper who adapts to a drifting quantum channel gains materially over one who does not, according to a paper newly posted to arXiv under the title 'Learnt Attacks on Quantum Key Distribution under Channel Noise and Device Drift'. The security analyses under which QKD links are provisioned assume stationary channels; the working channel is not stationary, and the paper supplies a number where none had been quantified. The authors pose adaptive eavesdropping as a constrained Markov decision process: the attacker selects one circuit per round as the noise level moves, with an abort condition over each block of rounds, and the value of adaptation is bounded between the best fixed circuit and a dynamic-programming upper bound. The headline figures are these. On device-independent E91 under bilateral depolarising noise, a reinforcement-learning attacker raises her Holevo information from 0.135 for the best fixed circuit to 0.348 at zero detection, 98 per cent of the upper bound. On BB84 under a drifting bit-flip channel, she exceeds a conservative noise-indexed rule by 0.024 in fidelity, 99 per cent of the upper bound. The search, started from random gate sequences, recovers the known analytical cloners and the collective-attack key rate, which is to say the method rediscovers attacks already established in the literature before finding new ones. Under stationary noise the attacker's gain from basis asymmetry changes sign depending on whether the error-rate constraint is averaged or per-basis; the same channel can reward opposite tactics under different accounting. The paper is a modelling result; it reports no incident in a deployed link, and none is claimed. The exposure lies in the provisioning assumption, which treats a moving channel as stationary, and in the interval between recalibrations, which is where the drift has its time to work. The attacker's method matters as much as her score. Where an earlier construction, that of Decker and colleagues, trained a parametrised circuit on a fixed gate template against a fixed channel, this search varies the gate structure and the rotation angles jointly, and returns circuits compact enough to form a discrete action set. That is what makes adaptation possible at all: the attacker holds a menu rather than a single instrument, and chooses from it as the channel drifts. The construction accordingly extends to noise models without a known template, the amplitude damping channel among them. In the model the noise level follows an Ornstein-Uhlenbeck process, and the abort condition is a budget applied over each block of rounds. The adaptive attacker is not assumed to alter the channel's own noise; she is assumed to watch it and choose accordingly. Under that constraint, the rise in Holevo information on E91 and the fidelity margin on BB84 are the paper's finding, not its premise. —Inspector Grey Dispatch from The Prepared E0