It might also be just correlation? Like, those agents were all instances of the same one or two models, so if that model has a preferred order it tries finding vulnerabilities in (the same way all current models have a particular writing style baked into them by RLHF), then most of the swarm will follow the same order and converge on the same services to exploit.